<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Monash DeepNeuron on Medium]]></title>
        <description><![CDATA[Stories by Monash DeepNeuron on Medium]]></description>
        <link>https://medium.com/@monashdeepneuron?source=rss-19e1ddc97f32------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*Q_jprls0EfhleMaXpE_r9g.jpeg</url>
            <title>Stories by Monash DeepNeuron on Medium</title>
            <link>https://medium.com/@monashdeepneuron?source=rss-19e1ddc97f32------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sat, 21 Nov 2026 17:57:10 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@monashdeepneuron/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[What do layoffs in the technology sector mean for the rest of us?]]></title>
            <link>https://monashdeepneuron.medium.com/what-do-layoffs-in-the-technology-sector-mean-for-the-rest-of-us-4c71c8461a3c?source=rss-19e1ddc97f32------2</link>
            <guid isPermaLink="false">https://medium.com/p/4c71c8461a3c</guid>
            <category><![CDATA[unemployment]]></category>
            <category><![CDATA[future-of-work]]></category>
            <category><![CDATA[tech-layoff]]></category>
            <category><![CDATA[fair-work-commission]]></category>
            <category><![CDATA[layoffs]]></category>
            <dc:creator><![CDATA[Monash DeepNeuron]]></dc:creator>
            <pubDate>Thu, 16 Apr 2026 20:49:57 GMT</pubDate>
            <atom:updated>2026-04-16T20:49:57.849Z</atom:updated>
            <content:encoded><![CDATA[<blockquote>By <strong>Adam Choong</strong> (MDN Law and Ethics Committee Member)</blockquote><p>Last month, Mike Cannon-Brookes, the co-founder of one of Australia’s tech startup darlings, Atlassian, announced that the software management company would be laying off 1600 jobs, approximately 10% of its workforce.¹ This comes amidst sharp selloffs of its stock in an increasingly volatile share market, resulting in more than half of the company’s valuation being wiped out since the start of the year. Just weeks before, ASX giant, WiseTech Global, also announced similar layoffs to 29% of its workforce, to occur over the next 18 months. The Software company claims this is in light of an AI transformation journey taking place.² WiseTech has also experienced similar valuation losses to Atlassian, seeing its stock plummet by almost 50% amid regulatory concerns being investigated by ASIC and raised by investors.³ While the technology sector is no stranger to extreme layoff events and sharp plunges in stock prices, with the emergence of more sophisticated generative AI capabilities, the most recent being Anthropic’s upgraded Claude models, the rest of the labour market is now anticipating alongside those in the technology sector for shocks to Australia’s workforce.</p><p>Within the banking sector, we’ve seen changes take place in the industry due to rapid AI acceleration, although some of these have fallen through the cracks. Most notably, the Commonwealth Bank of Australia late last year attempted to replace 45 roles in its call centres with automated voice bots but ultimately reversed this decision when employees complained of spikes in call-related workloads during the rollout.⁴ While this incident highlights shortcomings in current AI capabilities, the pace and scale of these technologies’ advancement are undeniable, with top ASX-listed companies continuing to invest heavily in partnerships with American technology giants such as Nvidia, AMD, AWS, OpenAI, Anthropic, and Microsoft. So, what do these incidents mean for the rest of us?</p><p>The Fair Work Commission faces its own challenges that might offer some clues. In a case set to be heard next year, which is likely to overlap with WiseTech Global’s and other Australian companies’ planned AI-related redundancies, Fair Work will decide whether employers must consult with employees before implementing significant workplace changes brought about by AI.⁵ While the Commission has historically been used to lessen the impact of change, in this case, the Chief Executive of the Australian Industry Group suggests that, unlike before, it’s improbable that this decision will alter how businesses operate, considering the rapid pace of AI development and its disruptive effects across various industries.</p><p>In response to these changes, to prepare those about to enter the workforce, universities are now making internships and work experience compulsory before graduation. embedding employability skills and networking opportunities throughout students’ degrees, rather than only at the end.⁶ Furthermore, not all sectors in Australia have been upended by the latest developments; healthcare, construction and social work still face significant shortages, meaning that AI isn’t coming for everyone’s jobs just yet.</p><p>The national unemployment rate as of February sits at 4.3%, and the RBA estimates that the jobless rate will hit 4.6% by mid-2026.</p><p>¹ Jaspreet Singh, ‘Atlassian to Cut Roughly 10% Jobs in Pivot to AI’ (online, 12 March 2026) &lt;https://www.reuters.com/technology/atlassian-lay-off-about-1600-people-pivot-ai-2026-03-11/&gt;.</p><p>² Stephanie Chalmers and Jasper Wells, ‘WiseTech Global to Cut 2,000 Jobs as AI Ends Era of “Manually Writing Code”’ (online, 25 February 2026) &lt;https://www.abc.net.au/news/2026-02-25/wisetech-job-losses-losing-2000-over-next-two-years-coding-era/106387486&gt;.</p><p>³ Max Mason and Kate McClymont, ‘Recording Reveals Richard White’s $5m Offer to WiseTech Complainant’ (online, 23 March 2026) &lt;https://www.afr.com/technology/recording-reveals-richard-white-s-negotiations-with-wisetech-complainant-20260318-p5oyk3&gt;.</p><p>⁴ Stephanie Chalmers, ‘Commonwealth Bank Backtracks on AI Job Cuts, Apologises for “error” as Call Volumes Rise’ (online, 21 August 2025) &lt;https://www.abc.net.au/news/2025-08-21/cba-backtracks-on-ai-job-cuts-as-chatbot-lifts-call-volumes/105679492&gt;.</p><p>⁵ David Martin-Guzman, ‘Surge in AI Dismissal Claims Delays AI Consultation Review’ (online, 13 March 2026) &lt;https://www.afr.com/work-and-careers/workplace/surge-in-ai-dismissal-claims-delays-ai-consultation-review-20260312-p5o9yr&gt;.</p><p>⁶ Melinda Ham, ‘How Unis Build Job-Ready Graduates from Day One’ (online, 12 February 2026) &lt;https://www.afr.com/work-and-careers/careers/how-unis-build-job-ready-graduates-from-day-one-20260128-p5nxph&gt;.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=4c71c8461a3c" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Balancing Innovation and Risk]]></title>
            <link>https://monashdeepneuron.medium.com/balancing-innovation-and-risk-fba2b67d3e6b?source=rss-19e1ddc97f32------2</link>
            <guid isPermaLink="false">https://medium.com/p/fba2b67d3e6b</guid>
            <category><![CDATA[ai-policy]]></category>
            <category><![CDATA[labour-market]]></category>
            <category><![CDATA[ai-regulation]]></category>
            <category><![CDATA[technology-ethics]]></category>
            <category><![CDATA[ai-governance]]></category>
            <dc:creator><![CDATA[Monash DeepNeuron]]></dc:creator>
            <pubDate>Mon, 23 Mar 2026 12:28:53 GMT</pubDate>
            <atom:updated>2026-03-23T12:28:53.735Z</atom:updated>
            <content:encoded><![CDATA[<h4><strong><em>The Productivity Commission’s Approach to AI Regulation</em></strong></h4><blockquote>By <strong>Eiman Yambio</strong> (MDN Law and Ethics Committee Member)</blockquote><p>Australia’s prolonged productivity stagnation has coincided with rapid advances in artificial intelligence (‘<strong>AI</strong>’). In August 2025, the Productivity Commission (‘<strong>PC</strong>’) released its interim report, <em>Harnessing Data and Digital Technologies</em>, which addresses three domains — AI, data sharing, and financial reporting infrastructure.¹</p><p>This article focuses on the Commission’s treatment of AI, where it estimates that machine intelligence could contribute more than $116 billion to the national economy within a decade and increase labour productivity by 4.3 percent.² The report situates Australia between two emerging regulatory models: the European Union’s risk-based legislative framework and the United Kingdom’s pro-innovation, principles-based approach.</p><p>For a country struggling to break the one percent productivity barrier, this is an enticing promise. But embedded in the Commission’s optimism lies a calculated policy trade off: innovation cannot be accelerated without eroding rights, widening inequality, or deepening digital divides.</p><p>The Commission’s guiding philosophy is unmistakably pro-growth. It calls on state and federal governments to remove regulatory friction, invest in digital infrastructure, and enable data sharing between firms, researchers, and public agencies. Yet, it also warns that excessive or premature regulation could “stifle opportunity.” The tension between risk and reward captures the report’s central message. Regulation should intervene only when necessary, and only once innovation has room to develop.³</p><h3><strong>I. A Productivity “First” Philosophy</strong></h3><p>The PC’s logic begins with an economic imperative. Australia’s productivity growth has stagnated for more than a decade, and the Report casts AI as a potential breakthrough comparable to the steam engine or electrification.⁴ The Commission situates AI not as a niche technology but as a general-purpose enabler capable of transforming healthcare, manufacturing, logistics, and public administration.</p><p>From this perspective, the question is not <em>whether</em> AI should be adopted, but how fast and under what conditions it should be. The Report’s preferred answer is an outcome-based regulatory philosophy. Existing frameworks — including privacy, competition, consumer protection, and discrimination — should be adapted to address AI related risks where necessary, while new AI specific legislation should be treated as a “last resort.”⁵ By advocating this position, the Commission aligns itself with a growing international bloc, led by the United States and the United Kingdom, that prioritises innovation leadership over precautionary control — a model most clearly articulated in the United Kingdom under a previously conservative government.⁶</p><p>It is a pragmatic argument, and one that resonates with industry. After all, Australia risks falling behind more agile economies if compliance costs outweigh investment incentives. However, the PC’s optimism reveals an underlying tension, it assumes that productivity and public trust can be pursued sequentially rather than simultaneously. In doing so, it leaves open the question of who bears the cost when innovation runs ahead of governance. By contrast, the European Union’s Regulation (EU) 2024/1689 of the European Parliament and of the Council (‘<strong>AI Act</strong>’) answers this directly: it allocates responsibility through ex ante risk tiers, prescriptive compliance duties and explicit accountability structures for high-risk systems.⁷ Rather than assuming that harms will be absorbed by markets or recovered later, the EU embeds cost-bearing obligations upfront — effectively socialising responsibility across developers, deployers and regulators. This stands in marked contrast to the PC’s sequencing logic, which implies that the distribution of risk can be worked out downstream.⁸</p><h3><strong>II. The Resistance to AI Specific Laws</strong></h3><p>The Commission’s resistance to bespoke AI legislation is rooted in economics. Regulation, it argues, imposes fixed costs that deter entry and experimentation, particularly for small firms.⁹ Moreover, because AI systems evolve rapidly, rigid rules could become obsolete before they are enacted. The PC instead promotes a principle-based model focused on outcomes, safety, transparency, accountability, rather than prescriptive inputs.¹⁰</p><p>This reasoning is sound in theory but fragile in practice. “Outcome-based” regulation often relies on self-assessment and voluntary standards, mechanisms that function most effectively here regulators possess strong oversight capacity. Comparatively research by the Organisation for Economic Co-operation and Development (<strong>‘OECD’</strong>) indicates that without adequate oversight and procedural legitimacy, such models may reduce compliance rather than enhance it.¹¹ Australia’s fragmented digital governance landscape, split between privacy regulators, consumer agencies, and competition authorities, makes such coherence difficult.¹² Without stronger coordination, the promise of flexible governance can easily collapse into regulatory ambiguity. A coordinated model could involve a central lead agency or cross-regulatory taskforce empowered to set standards, allocate jurisdiction and coordinate enforcement. The Australian Competition and Consumer Commission’s (‘<strong>ACCC</strong>’) successful s 18 action against Microsoft for misleading claims around Co-pilot adoption demonstrates both the effectiveness and the limitations of Australia’s current regulatory approach.¹³ The proceedings were resolved swiftly through settlement, signalling strong consumer protection enforcement. However, the conduct engaged competition law, consumer protection, corporate governance and intellectual property (‘<strong>IP</strong>’) concerns, yet was pursued through a single statutory pathway. In a coordinated environment, regulators would share data, assign responsibility and develop consistent compliance expectations, rather than responding in siloed bursts.</p><p>The Commission’s light-touch stance also reveals its economic priorities. By placing innovation ahead of precaution, it mirrors the logic of what might be called the AI productivity gamble: betting that the social dividends of acceleration will outweigh the risks of disruption. This approach is not reckless; indeed, it reflects the Commission’s mandate to improve efficiency and growth, but it sidelines the broader human consequences that lie beyond the balance sheets.</p><h3>III. The Human Gap</h3><p>While the PC acknowledges that AI adoption will cause “painful transitions” in the labour market, its policy response remains underdeveloped.¹⁴ The report gestures toward reskilling programs and workforce mobility but stops short of addressing systemic inequality or the psychosocial cost of technological displacement. In its calculus, productivity gains appear as national aggregates, not as lived experiences.</p><p>This gap matters because the benefits of AI will not be evenly distributed. The Commission assumes that displaced workers will find new roles in higher value sectors, but empirical evidence suggests otherwise. Recent research from the Centre for Future Work at The Australia Institute similarly cautions that productivity gains are not evenly distributed; structural barriers, insecure employment and unequal wage pass-through mean lower-paid and precarious workers often experience technological change as exclusion rather than mobility.¹⁵ Without targeted safety nets, the “transition” may feel less like mobility and more like exclusion.</p><p>A similar gap appears in the creative industries. One of the most contentious elements of the PC’s report is its proposal to explore a text and data mining (‘<strong>TDM</strong>’) exception to Australian copyright law, allowing AI systems to scrape and train on creative works without prior permission. The Commission frames this as essential for innovation and competitive parity with jurisdictions such as the United States and Japan.¹⁶ Yet for authors, musicians, and visual artists, it signals the erosion of both economic and moral rights.</p><p>Creative Australia’s submission to the Commission raised this concern directly, warning that prioritising broad data access rights for AI developers over creator consent could undermine the cultural economy and weaken Australia’s creative industries.¹⁷ In other words, the PC’s productivity lens may measure what AI contributes to GDP, but not what it subtracts from creative livelihoods.</p><h3><strong>IV. Balancing Innovation and Rights</strong></h3><p>None of this denies the PC’s core insight that overregulation can suffocate technological dynamism. Yet the tension between innovation and regulation is often overstated. Governance does not inevitably slow growth, it can build trust, stability and diffusion — conditions that make productivity gains possible.</p><p>A more balanced model would treat AI governance as a dual mandate: promoting innovation and protecting human dignity. That means embedding rights-based principles — transparency, fairness, explainability — within outcome-based frameworks instead of treating them as afterthoughts. It also requires clearer lines of institutional responsibility. Who ensures that algorithmic decisions are auditable? Who verifies that public sector AI complies with administrative law values such as procedural fairness? These questions engage core principles of Australian administrative law, including procedural fairness, accountability, and reviewability.</p><p>Other jurisdictions offer instructive contrasts. The EU’s <em>AI Act</em>, for all its bureaucratic sprawl, at least defines risk categories and allocates compliance duties.¹⁸ The United Kingdom’s pro-innovation strategy, by contrast, relies on cross regulator coordination and voluntary standards, an approach the PC appears to emulate. Australia could synthesise both models, combining the EU’s clarity with the UK’s agility, to craft a governance architecture that supports innovation without abdicating responsibility.¹⁹</p><h3><strong>V. Toward Smarter Governance</strong></h3><p>The PC’s interim report is an ambitious blueprint for economic renewal, but its success will depend on whether governments translate aspiration into credible safeguards. Productivity and protection are not mutually exclusive; they are mutually reinforcing. A workforce that trusts technology is more likely to adopt it. Consumers who understand algorithmic decision making are more likely to engage with it.</p><p>A truly “outcome-based” regulatory framework should therefore define whose outcomes matter. Economic efficiency is one outcome; social inclusion, creative integrity, and equitable opportunity are others. A narrow focus on GDP gains risks transforming AI into a winner-takes-most game, one that boosts output but erodes cohesion.</p><p>To avoid that fate, policymakers must move beyond the productivity lens. They should view AI not merely as a set of tools but as a sociotechnical system that reorganises work, culture, and governance itself. That requires investing in digital rights literacy, worker transition programs, and ethical design capacity across the public sector. These are not constraints on innovation; they are the infrastructure of sustainable growth.</p><h3>VI. Conclusion</h3><p>The Productivity Commission’s AI strategy is a calculated bet on the future: that by holding back from heavy handed regulation, Australia can unleash a new wave of productivity. The wager may pay off, but only if the country also invests in the human and ethical foundations that make innovation worth pursuing.</p><p>AI will reshape the economy whether governments are ready or not. The question is whether Australia’s policy institutions can shape that transformation with foresight rather than hindsight. If the Commission’s optimism is to become more than rhetoric, its productivity blueprint must evolve into a governance model that measures success not just by how efficiently machines learn, but by how wisely humans lead.</p><p>¹ Productivity Commission, <em>Harnessing Data and Digital Technologies </em>(Interim Report, August 2025).</p><p>² Ibid 100.</p><p>³ Ibid 5.</p><p>⁴ Ibid 6.</p><p>⁵ Ibid 22.</p><p>⁶ UK Government, <em>A Pro-Innovation Approach to AI Regulation </em>(CP 815, 2023) 1.1.20; White House Office of Science and Technology Policy, Blueprint for an AI Bill of Rights: Making Automated Systems Work for the American People (White Paper, October 2022).</p><p>⁷ Regulation (EU) 2024/1689, arts 6, 9–15; Annex III.</p><p>⁸ Productivity Commission, <em>Harnessing Data and Digital Technologies </em>(Interim Report, August 2025) 22–24.</p><p>⁹ Ibid 100.</p><p>¹⁰ David Ruountree, Tracey Lu, Tom Hall, Emily Cravigan, “Harnessing AI and data through a new regulatory blueprint? Productivity Commission diverts the conversation” (2025) &lt; <a href="https://www.allens.com.au/insights-news/insights/2025/08/Harnessing-AI-and-data-through-a-new-regulatory-blueprint-Productivity-Commission-diverts-the-conversation/">https://www.allens.com.au/insights-news/insights/2025/08/Harnessing-AI-and-data-through-a-new-regulatory-blueprint-Productivity-Commission-diverts-the-conversation/</a> &gt;.</p><p>¹¹ OECD, Regulatory Policy Outlook 2021 (Report, 2021) 187–188.</p><p>¹² Attorney-General’s Department, <em>Privacy Act Review Report</em> (Report, February 2023) 70, 270.</p><p>¹³ Australian Competition and Consumer Commission<em>,</em> ‘Concise Statement’, Submission in <em>Australian Competition and Consumer Commission v Microsoft Pty Ltd</em>, NSD1977/2025, 27 October 2025.</p><p>¹⁴ Productivity Commission, <em>Harnessing Data and Digital Technologies </em>(Interim Report, 2025) 13.</p><p>¹⁵ Jim Stanford, ‘<em>Productivity in the Real World: What it is, what it isn’t, and how to make it work better for worker’ </em>(Research Paper,<em> </em>The Australia Institute Centre for Future Work, July 2025) 51–62.</p><p>¹⁶ Productivity Commission, <em>Harnessing Data and Digital Technologies </em>(Interim Report, 2025) 27.</p><p>¹⁷ Creative Australia, <em>Submission to the Productivity Commission, Harnessing Data and Digital Technologies </em>(15 September 2025) 4–5.</p><p>¹⁸ Regulation (EU) 2024/1689.</p><p>¹⁹ UK Government, <em>A Pro-Innovation Approach to AI Regulation</em> (CP 815, 2025) Annex A.1.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=fba2b67d3e6b" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[What Is Outreach at Monash DeepNeuron?]]></title>
            <link>https://monashdeepneuron.medium.com/what-is-outreach-at-monash-deepneuron-4a391cf84529?source=rss-19e1ddc97f32------2</link>
            <guid isPermaLink="false">https://medium.com/p/4a391cf84529</guid>
            <category><![CDATA[engineering]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[education-day]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Monash DeepNeuron]]></dc:creator>
            <pubDate>Fri, 24 Jan 2025 07:44:30 GMT</pubDate>
            <atom:updated>2025-01-24T07:44:30.414Z</atom:updated>
            <content:encoded><![CDATA[<h4>Spark &amp; Share <em>(Igniting minds with insights and collective learning)</em></h4><figure><img alt="(Collective word data visualisation on how members describe Outreach)" src="https://cdn-images-1.medium.com/max/1024/1*Y7EaVSrxXryeKEjZtZZV-w.png" /><figcaption>Collective data visualization of members’ perspectives on the meaning of ‘Outreach’.</figcaption></figure><p>Outreach has developed over the years into more than just one branch. What started as a single sub-branch has grown into a multi-faceted platform. Today, Outreach encompasses distinct committees, including Law and Ethics, a dedicated Outreach branch within Education, and Outreach Ancillary, which draws contributors from various other branches.</p><p>Greater Outreach now serves as a space for STEM students to explore their passions and embrace the art of education. Reflecting on this evolution, Jayus Kion (co-Education lead) highlights, “Cultivating this initiative has been the most significant achievement of my career in DeepNeuron. I am so proud of the integrity and excellence we have brought into high school and university education.”</p><h3>At a quick glance…</h3><p>A major turning point for the Outreach team’s growth also involved a shift in focus on how workshops are presented. As Andrique (Outreach member) explains, “Rather than trying to inscribe everything we know about AI into workshop slides, Outreach has matured to consider what information is important for different audiences and how we can adapt our teaching styles to make our workshops more digestible and valuable for them.”</p><p>In continuation, another key part of Outreach’s transformation lies in the personal touch its members bring to workshops. As Sanithma (Outreach member) tells us, “This year, the teaching style in MDN Outreach has drastically evolved. Workshops have become increasingly interactive, actively involving students in the learning process and immersing them in the fascinating world of AI. Despite the inherently technical nature of the topics, we’ve embraced a light-hearted approach, blending entertaining jokes and engaging activities with the core goal of educating students on the fundamental theories of AI.”</p><h3>Marketing Gets Up Close with Outreach: Highlights from the Team</h3><h4>We had the chance to chat with some incredible members of our Outreach team about their experiences, the workshops they’ve been part of, and the moments that have stood out to them. Here’s what they had to share:</h4><p><em>A highlight for me was presenting our Semester 1 ENG1013 ‘Ethics of AI’ presentation and seeing the way students actively engaged with the content. Having completed the same unit a year prior it was exciting to be on the other side of the Workshop and hear first year perspectives on the intersection between AI, engineering and the associated ethical implications.</em></p><ul><li><em>Olivia Pepe, Law and Ethics Co-Manager</em></li></ul><p><em>Visiting the Landers and Rogers office and seeing the view truly inspired me to pursue a career in law. Being part of this team allows me to educate and empower students, showing them that AI isn’t as intimidating as it seems and emphasizing that the responsibility lies with them to train reliable and transparent AI models.</em></p><ul><li><em>Wendy Huang, Law and Ethics Awareness + Outreach (AO) Member</em></li></ul><p><em>I’ve taught the Introduction to HPC series, which was a lot of fun, and really put me out there. I truly believe that for some people I was able to break down the beliefs of HPC being super difficult and hard to enter to something a lot simpler to everyone inherently uses. I’ve also done , ENG1013 Intro to AI, Workshop Workshop and Intro to OOP for highschool kids. Another highlight for me was the Engage Engineer workshop that gave highschool kids a little intro into being a software engineer. This highlight stands out to me because of how amazing the kids were, as they often came in not caring about software engineering but came out with an interest in software engineering and also AI.</em></p><ul><li><em>Jackie Nguyen, HPC Co-Lead</em></li></ul><p><em>Being a part of outreach means closing the gap on AI education (or just education in general) at any level and for every demographic. Outreach is building as many bridges to navigate the next generations into building confidence and comfortability of the enterprises of Artificial Intelligence. Where there is an opportunity to learn or teach AI, we want to be there!</em></p><ul><li><em>Jayus Kion, Education Co-Lead</em></li></ul><p><em>One of the major highlights of teaching is witnessing that “Aha!” moment on a student’s face. It’s absolutely priceless to see someone transition from confusion to understanding — from “what on earth is this?” to “I actually get it now.” That’s exactly why I teach, and I know it resonates with many of us in this profession.</em></p><ul><li><em>Sarge, Education Co-Lead</em></li></ul><p><em>I have designed and taught workshops for events such as VicHack, Robogals Hackathon and Engage Engineer, collaborations with MAC, CCA and MYMI, as well as school workshops with Scotch College, WCI and so on. The most rewarding part was attending Outreach meetings at the end of the year and realising how close I have gotten to my team as a result :D</em></p><ul><li><em>Andrique H, Education Outreach member</em></li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/426/1*WyWhMFAJpZ7V7xA5MIJr3Q.gif" /><figcaption>Slide taken from Outreach Robogals Workshop</figcaption></figure><p><em>My main highlight is the creative freedom we have when it comes to helping design certain educational content. I absolutely love when we get to experiment with different visual elements like 3d models, and engaging analogies that really challenge the way AI and HPC education is delivered.</em></p><ul><li><em>Sanithma M, Education Outreach Member</em></li></ul><p>Looking to the future, there’s an overwhelming sense of excitement and ambition within the team. Members envision an evolving and expanding Education branch, with aspirations to reach more schools, extend university workshops to additional units, and foster greater collaborations. The focus is on delivering higher-quality and more impactful workshops, events, and educational experiences. From whimsical animations to wider outreach, the team is committed to seizing every opportunity to educate and inspire, ensuring that what lies ahead will be nothing short of extraordinary.</p><p><em>MDN Outreach is not just a platform for education; it’s a movement that challenges norms and empowers students. By blending creativity, passion, and forward-thinking, the team is reshaping how students connect with AI and inspiring the innovators of tomorrow.</em></p><p><strong><em>Special thanks to:</em></strong></p><p><strong><em>Publication editor: Regina Lu</em></strong></p><p><strong><em>Publication assistants: Quynh Do, Sanithma M, Veronica He</em></strong></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=4a391cf84529" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Library Artefact Digitisation AI Project — Blog Post Q&A]]></title>
            <link>https://monashdeepneuron.medium.com/library-artefact-digitisation-ai-project-blog-post-q-a-cc4ec75ccfc8?source=rss-19e1ddc97f32------2</link>
            <guid isPermaLink="false">https://medium.com/p/cc4ec75ccfc8</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[swe]]></category>
            <category><![CDATA[computer-science]]></category>
            <category><![CDATA[deep-learning]]></category>
            <category><![CDATA[hpc]]></category>
            <dc:creator><![CDATA[Monash DeepNeuron]]></dc:creator>
            <pubDate>Tue, 28 May 2024 23:27:22 GMT</pubDate>
            <atom:updated>2024-05-28T23:27:22.644Z</atom:updated>
            <content:encoded><![CDATA[<h3>Library Artefact Digitisation AI Project — Blog Post Q&amp;A</h3><h4>Deep dive into an AI project — featuring Pranav Kutty and Ryan Garg</h4><p><strong>Featuring Authors:</strong></p><p><em>Pranav Kutty — Project Manager</em></p><p><em>Ryan Garg — Project Member</em></p><p>We had the opportunity to hear from Pranav Kutty, current project manager for the Library Artefact Digitisation project and Ryan Garg, who is a project member on this project. Both Pranav and Ryan share their insights from this project which involves the digitisation of real-world objects into viewable/interactive three dimensional models! This project is a collaboration with Monash Automation, another student engineering team, who are working on the creation of robots that can take videos and images of objects which will be digitised.</p><p><strong>Can you tell me a little bit more about that project? As well as your aspect of the project collaboration?</strong></p><p><em>Pranav </em>— At the beginning of this project, the idea was that we would work with the Sir Louis Matheson Library to digitise some of the artefacts in their collection, so that they could have digital versions available on their website for the public to view. Since then, our project has evolved into pursuing the idea that we can create a framework to digitise any object. For those unfamiliar with<strong> digitisation</strong>, as the name suggests, it is the process of converting items into a digital form. In our context, we aim to convert objects into a digital 3D image, allowing users to scroll around and view all angles in a digital format as if they had the object in front of them. A great example of this is the George Washington artefact digitisation from the Smithsonian Museum, linked below.</p><p>This involves taking images or videos of the object and then training models on this data to produce a final digitised output. This project is in collaboration with Monash Automation, who are currently working on a robot that can take videos and images of the objects we want to digitise. We have been exploring the various existing digitisation processes and comparing their outputs after training to determine what works best for us.</p><p>As team lead, my role in this project is primarily centred around planning, organisation, and communication, as well as contributing to model training and research.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*XZw8_pA-Fq2pVRS7sV-JBw.gif" /><figcaption>Image Source: Smithsonian Institution, “George Washington,” 3D Digitization, <a href="https://3d.si.edu/object/3d/george-washington%3A789cf90a-4387-4ac1-9e96-c7d6a7b9d26f"><strong>[Click here for more]</strong></a></figcaption></figure><p><strong>What technologies and methodologies are being employed in the development of this AI project?</strong></p><p><em>Both</em> — This project thus far has centered around two technologies: <a href="https://www.youtube.com/watch?v=CRlN-cYFxTk"><strong>Neural Radiance Fields (NeRFs)</strong></a> and <a href="https://www.youtube.com/watch?v=sQcrZHvrEnU"><strong>Gaussian Splatting</strong></a>. Both of these are machine learning methods of digitisation, and both are installed and run predominantly using the command line. At its essence, a NeRF functions as a neural network generating a 3-dimensional scene representation from 2-dimensional input images. However, within the realm of NeRFs, numerous variants exist, and our current focus lies in exploring these diverse implementations. Some of the NeRF variants we are delving into include Instant-NGP, Mip-NeRF, and PyNeRF.</p><p>Moving forward, we will be shifting our focus to the different models of Gaussian Splatting. The plan is to explore all available methods thoroughly, and once we have done so, we aim to develop our own model. By experimenting with and understanding the various approaches within NeRFs and Gaussian Splatting, we hope to contribute innovative solutions to the field of machine learning-based digitisation.</p><p><strong>How do you envision this AI project making a positive impact or addressing real-world issues?</strong></p><p><em>Ryan </em>— Whilst at the moment we’re working on digitising arbitrary objects at Monash University, the technology of creating 3-dimensional scenes from sets of 2-dimensional images can have many real-world applications. For instance, there are studies being conducted on the effectiveness of rendering medical imagery such as MRI and ultrasound scans. Additionally, there are many applications in aerial imagery, and NeRFs can be used in professions such as urban planning.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/600/1*Y3j2aGOES3RTY0x8Rbg98g.gif" /><figcaption>Machine learning digitization based demo of dragon object (work in progress)</figcaption></figure><p><em>Pranav</em> — Given that we are attempting to produce an output for Matheson Library, and that they intend our work to be used on their website, it is fair to say that this project does have a tangible real world impact. I think the positive impact that this project has stems from the fact that it makes these artifacts and objects accessible without having to see them in person. It allows one to understand what something might look like in person, in a way that a simple 2D image would not be able to.</p><p><strong>What were some key lessons learned by the team/or that you learnt from collaborating on this project?</strong></p><p><em>Pranav —</em> This being a new project and my first time as project manager, it was a massive learning curve. In our first meeting as a team we had a brief discussion about what we knew about object or artifact digitization and not one person had any background. As team lead this was a scary experience, given that I was expected to plan a project about something that I had never seen before. Thankfully during our meetings the team was able to help me with planning our next steps at each stage by voicing their opinions. The key lesson that I learned in that scenario was mainly to trust the people in my team and their judgement.</p><p><em>Ryan —</em> A key lesson we learnt in this project is that there is no single correct technique or model to use for our task. Through our experimentation and research, we discovered that every model has its advantages and disadvantages, and it depends on the object we are attempting to digitise. We also learnt the importance of understanding the theory behind the methods we are using, as it makes it significantly easier to understand why a method may not be particularly effective for a specific use case, or why it might perform better than other methods.</p><p><strong>Could you please give me three words to describe this project?</strong></p><blockquote><em>Pranav </em>— Fresh, rewarding and ambitious.</blockquote><blockquote><em>Ryan</em> — Innovative, challenging and multifaceted.</blockquote><p><strong><em>Interested in learning more?</em></strong></p><p>Check out these external links:</p><p><a href="https://3d.si.edu/object/3d/george-washington%3A789cf90a-4387-4ac1-9e96-c7d6a7b9d26f">3D Digitisation example of George Washington </a>— credit towards the Smithsonian National Museum of American History</p><p><a href="https://www.youtube.com/watch?v=CRlN-cYFxTk">NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis (ML Research Paper Explained)</a></p><p><a href="https://www.youtube.com/watch?v=sQcrZHvrEnU">3D Gaussian Splatting — Explained!</a></p><p><strong>AI Project Library Artefact Digitization current team:</strong></p><p><em>Pranav Kutty, Ryan Garg, Lily Ung, Laura White</em></p><p><strong>Special thanks to our editors: </strong><em>Reya Jain and Regina Lu</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=cc4ec75ccfc8" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[A Law and Ethics Blog: “AI, EU Ready Kids?”]]></title>
            <link>https://monashdeepneuron.medium.com/a-law-and-ethics-blog-ai-eu-ready-kids-684be1de9952?source=rss-19e1ddc97f32------2</link>
            <guid isPermaLink="false">https://medium.com/p/684be1de9952</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[ethics-in-tech]]></category>
            <category><![CDATA[ai-regulation]]></category>
            <category><![CDATA[law]]></category>
            <category><![CDATA[european-law]]></category>
            <dc:creator><![CDATA[Monash DeepNeuron]]></dc:creator>
            <pubDate>Wed, 22 May 2024 00:02:09 GMT</pubDate>
            <atom:updated>2024-05-22T07:44:32.185Z</atom:updated>
            <content:encoded><![CDATA[<blockquote>Featured Author: Flynn Gill</blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*IU6NjAlNHBNF8Jg1" /><figcaption>Photo by <a href="https://unsplash.com/@nasa?utm_source=medium&amp;utm_medium=referral">NASA</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p><strong><em>1 AI’s Global Regulation</em></strong></p><h4>The ‘European AI Act’ is fast approaching,¹ and now that the text of the ‘Act’ has been finalised this article will elucidate its content and explore why it has been awaited with cautious anticipation across the globe.</h4><p>It is not unusual for the European Union to indirectly set global standards through the unilateral adoption of European law by outside nations.² With a population of around 448 million people in the European Union,³ it is arguable that tremendous market pressure and regulatory influence will induce the Big Tech giants to capitulate to regulation of their AI system technology.⁴ This is a phenomenon that has been observed across many EU regulations, described as the ‘Brussels Effect’.⁵</p><p>However, care should be taken in assuming that a ‘Brussels Effect’ will occur for the ‘European AI Act’.⁶ Unlike previous regulations such as the General Data Protection Regulation, the ‘AI Act’ is deliberately interwoven with an extensive list of other regulations that have been designed to operate holistically to achieve ‘Europe’s Digital Decade’.⁷ Therefore, it is difficult to extract the ‘AI Act’ out in isolation without creating statutory interpretation issues.</p><p>It is possible that Australia may follow the Council of Europe’s newly adopted AI Convention instead,⁸ or other jurisdictions’ attempts — such as President Biden’s Executive Order concerning AI.⁹ Nonetheless, it remains an important regulation to have at least a rudimentary understanding of.</p><p><strong>2 <em>The ‘European AI Act’</em></strong></p><p>Contrary to its colloquial name, the ‘European AI Act’ is not an ‘Act’ at all. It is a proposed regulation from the European Union. The term regulation for this jurisdiction is hard law, however, and is as binding as any Australian Act would be in Australia. As is the nature of any law, ironically this ‘AI Act’ is not intelligent. It is not intelligent in the sense that AI is undergoing an exceptionally fast rate of innovation and technological change, and yet the law is not itself able to adapt easily. This has been problematic for the drafting of this particular regulation, as it went through transformative amendments in June 2023 after the explosion in popularity of generative AI.¹⁰</p><p><strong><em>a) Definition of Artificial Intelligence</em></strong></p><p>The regulation has defined what artificial intelligence is. While seemingly simple, there is no uniform international definition of AI.¹¹ Australia specifically has no legislative definition of AI. Any sort of stance on a definition can only be pieced together by various government websites and policy reports, all of which have a different definition.</p><p><strong><em>The finalised definition in the AI regulation is as follows: </em></strong>‘</p><p>AI system’ means a machine-based system that is designed to operate with varying levels of autonomy and that may exhibit adaptiveness after deployment, and that, for explicit or implicit objectives, infers, from the input it receives, how to generate outputs such as predictions, content, recommendations, or decisions that can influence physical or virtual environments’.¹²</p><p>This definition is adopting the OECD’s new updated definition of AI.¹³ The OECD is the most influential non-legal international organisation of its kind, frequently making recommendations that have been taken up by many jurisdictions across the world. Given the OECD’s widespread influence, further propagating this new definition encourages nations to agree upon a uniform definition, which can be viewed as a positive step towards eliminating cross-jurisdictional confusion. Australia may wish to consider supporting this goal by adopting a similar definition of AI.</p><p><strong><em>b) So, What Is It?</em></strong></p><p>The draft regulation is a market regulating instrument that will have a rollout effect of two years from the date of publication for many of its rules.¹⁴ Because of widespread commentary and attention, the regulation has been coaxed in discussions of protecting human rights and values.¹⁵ This discussion has resulted in some injections of human values into the regulation post the 2021 draft version, however the regulation more accurately reflects the European Commission’s familiarity with product safety legislation.¹⁶</p><p>Ultimately, despite the recital to the regulation containing references to the upholding of human rights and fostering AI innovation, its purpose is neither of these. Instead, its purpose is to assure European citizens that the AI system products that they are using are safe.¹⁷</p><p><strong><em>c) What is the Content of this Regulation?</em></strong></p><p>In achieving its purpose of facilitating the proper functioning of the European market, it has created harmonised rules on the development of AI.¹⁸ Further showing its alignment with product regulation as opposed to human rights legislation, the regulation has taken a risk-based approach, not a rights-based one.¹⁹ It is important to note that this regulation covers a tiny sliver of all AI technology, but of the AI systems it does cover the European Union has broken them down into four categories of varying risk:</p><ol><li>Unacceptable risks;</li><li>high risks;</li><li>limited risks; and</li><li>minimal risks.²⁰</li></ol><p>In doing so the regulation recognizes that some AI systems are innocuous, while others can pose a serious risk of harm. To grasp how this would play out in practice, it is useful to explore examples of AI in the different categories of risk.</p><p><strong><em>(i) Unacceptable Risk AI</em></strong></p><p>One clear example of AI technology in the unacceptable risk category is social scoring technology. This AI attributes a social score to individuals within the population based on key defined social factors such as occupation and residential living, as well as ingrained social biases.</p><p>This score is used to assess an individual’s ‘trustworthiness’ for the purposes of many essential decisions, such as banks lending money or whether insurance is granted. Given AI systems are already vulnerable to perpetuating biases based on discriminatory inputs, this type of AI system is particularly susceptible to being utilised as a tool to further perpetuate inequality and injustice. These unacceptable risk technologies are therefore banned outright.²¹</p><p><strong><em>(ii) High Risk AI</em></strong></p><p>Most of the regulation is centred on controlling this high risk category.²² High risk technology describes AI systems that have potential for human rights abuse if left unregulated. For this reason, the regulation sets out varying degrees of requirements for the development of these types of technology.</p><p>Annex III lists examples of high risk AI systems.²³ One of the examples listed is the use of AI for law enforcement. However it is relevant to note that since an AI system can be used for a wide spectrum of tasks, the regulation is filled with exceptions, and this is especially evident when this type of technology is used for military or national security purposes.²⁴</p><p><strong><em>(iii) Limited Risk AI</em></strong></p><p>These technologies also have the ability to breach human rights but are seen as a step down in threat and likelihood of harm compared to the high risk systems. Chatbots are a common example of AI systems in this category. Limited risk AI systems are subject to transparency obligations which can require that users are informed when they are interacting with an AI system.²⁵</p><p><strong><em>(iv) Minimal Risk AI</em></strong></p><p>Minimal risk AI are AI systems that are so innocuous that voluntary codes of conduct are seen as adequate to ensuring that the technology is used in a safe way.²⁶ AI used for inventory management or video games are a clear example of these relatively harmless AI technologies.</p><h4><strong><em>3 Conclusion</em></strong></h4><h4>The ‘European AI Act’ is an EU regulation designed to regulate the AI technology market. It does this by adopting a risk-based approach of product control, rather than a rights-based one for European citizens. There will be a rolling out period of two years for much of the regulation, and so the effects will be felt years from now.</h4><h4>The regulation primarily focuses on what it classifies as ‘high risk AI’, providing extensive regulation of this kind of technology. However, the AI systems it does cover is a niche of all total AI technology. In totality, the regulation is one piece in a network of EU regulations pre-existing and yet to come, and it will need to be read in conjunction with these other regulations.</h4><p>— — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — —</p><p>¹Regulation (EU) 2024/138 of the European Parliament and of the Council laying down harmonised rules on artificial intelligence and amending Regulations [2024] Official Journal TA (‘EU AI Act’).</p><p>²Anu Bradford, ‘The Brussels Effect: How the European Union Rules the World’ (2020) Oxford University Press, 25–26 (‘The Brussels Effect’).</p><p>³‘Facts and Figures on Life in the European Union’, European Union (Webpage) .</p><p>⁴The Brussels Effect (n 2) 26–36.</p><p>⁵Ibid 2–4.</p><p>⁶Chris Marsden and George Christou, ‘What Europe’s AI Regulation Moment Will Mean for the World’ (2 August 2023) 360info, [15]-[17].</p><p>⁷Monash Law, ‘EU AI Act A Trustworthy Framework that Respects Human Rights’ (YouTube, 20 March 2024), 00:11:27–00:12:17 .</p><p>⁸‘Council of Europe Adopts First International Treaty on Artificial Intelligence’, Council of Europe (Webpage, 17 May 2024) .</p><p>⁹88 FR 75191 — Safe, Secure, and Trustworthy Development and Use of Artificial Intelligence (2023) Federal Register 88(210) 75191, 75193.</p><p>¹⁰EU AI Act (n 1) ch V.</p><p>¹¹Giusella Finocchiaro, ‘The regulation of artificial intelligence’ (2023) AI &amp; Soc, 1–2.</p><p>¹²EU AI Act (n 1) art 3.</p><p>¹³OECD, ‘Explanatory memorandum on the updated OECD definition of an AI system’ (Research Report No 8, OECD Publishing, 2024), 4.</p><p>¹⁴ EU AI Act (n 1) 157 [179].</p><p>¹⁵ Jean Monnet Centre of Excellence JUST-AI, ‘JUST-AI JMCE ‘Meet the Expert Podcast’ n°2 — ‘The AI Act: Behind the scenes’ with Luca Bertuzzi’ (YouTube, 23 February 2024), 00:22:22–00:22:42 (‘Meet the Expert Podcast’). ¹⁶ Ibid 00:07:50–00:08:45, 00:12:26–00:12:45.</p><p>¹⁷ Ibid 00:22:22–00:24:59.</p><p>¹⁸ EU AI Act (n 1) 7 [8].</p><p>¹⁹ Meet the Expert Podcast (n 15) 00:12:26–00:12:45.</p><p>²⁰ ‘High-Level Summary of the AI Act’, EU Artificial Intelligence Act (Webpage, 27 February 2024) .</p><p>²¹ EU AI Act (n 1) ch II art 5.</p><p>²²Ibid ch III.</p><p>²³Ibid Annex III.</p><p>²⁴Ibid art 2.</p><p>²⁵Ibid ch IV.</p><p>²⁶ Ibid 148 [165].</p><p>— — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — —</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=684be1de9952" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[HPCxAI: Neural Cellular Automata Q&A]]></title>
            <link>https://monashdeepneuron.medium.com/hpcxai-neural-cellular-automata-q-a-e4e9a5f0e8a5?source=rss-19e1ddc97f32------2</link>
            <guid isPermaLink="false">https://medium.com/p/e4e9a5f0e8a5</guid>
            <category><![CDATA[hpc]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Monash DeepNeuron]]></dc:creator>
            <pubDate>Sun, 12 May 2024 11:44:33 GMT</pubDate>
            <atom:updated>2024-08-05T00:54:41.536Z</atom:updated>
            <content:encoded><![CDATA[<h3>Neural Cellular Automata Blog Post Q&amp;A</h3><h4>A HPC and AI collaboration — featuring Nyan Kyaw (AI Branch) and Keren Collins (HPC Branch)</h4><p><strong>Featuring Authors:</strong></p><p><em>Nyan Kyaw — AI Assistant Lead, AI Project Manager and Project Team Member for Cellular Automata</em></p><p><em>Keren Collins — HPC Project Manager and Project Team Member for Cellular Automata</em></p><p>We had the opportunity to hear from Keren Collins, current HPC member and project manager for Neural Cellular Automata and Nyan Kyaw, AI Assistant Lead and project team member. Both Keren and Nyan share their insights from this project which is a collaboration between our High Performance Computing (HPC) and Artificial Intelligence (AI) branches!</p><p><strong>Can you tell me a little bit more about that project? As well as your aspect of the project collaboration?</strong></p><p>Nyan (AI)— If you have ever heard of <a href="https://www.youtube.com/watch?v=xP5-iIeKXE8"><strong>Conway’s Game of Life</strong></a>, it’s essentially that. Cellular automata are essentially little isolated cells which follow a rule set that you determine.</p><p>In our case, we’re exploring a fun application using emojis. We input an emoji into our neural network, which then attempts to learn a rule set, gradually transforming our game from a single black pixel to that emoji. Currently, we’re in the process of transitioning it onto the web for a neat demo on our website. We already have functional models capable of generating emojis from a single black pixel.</p><p>Keren (HPC) — In this project there are two core components: the simulation grid, which defines the interactions of the cells and how they ‘perceive’ themselves and other cells, and a machine learning algorithm which defines how these cells behave/update themselves in each cycle based on this information. The HPC side of the project works on the simulation grid. We also work on how to allow these simulations to run smoothly and efficiently locally on the computer of any person who comes across the website we are developing.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*rD3KJBOSelyhiiWsuRsf8A.gif" /><figcaption>The model has acquired the capability to replicate the butterfly patterns in a manner such that when subjected to distortion (for instance, by moving the cursor over the pattern to “eradicate” portions of it), it adeptly learns to restore the original pattern.</figcaption></figure><p><strong>What value do you believe this project holds currently/in future?</strong></p><p>Nyan (AI) —In our current work, while we’re primarily replicating existing literature, we’re also excited about delving into <strong>entropy’s</strong> role in pattern learning. <strong>Entropy</strong>, in this context, reflects the evolving information content of patterns over time. Our innovative approach involves harnessing entropy to train neural networks dynamically, promising both novelty and research significance. However, significant value is also emerging from our HPC efforts. The HPC team is dedicated to parallelizing our Neural Cellular Automata application, aiming for exceptional speed and efficiency in cell calculations. Their outstanding work deserves a shoutout!</p><p><strong>Passing it over to the HPC team! What hardware and software resources did HPC employ for the project, given the focus on parallel processing?</strong></p><p>Keren (HPC) — Currently, since our objective is to make everything that we develop web-based, the HPC side has been working with WebGPU, based in JavaScript. We use WebGPU so we can utilise the parallel processing capabilities of the viewer’s PC to run our website’s interactive simulations.</p><p><strong>What were the considerations behind these choices?</strong></p><p>Keren (HPC)— Our main concern from the HPC point of view is to ensure as many of the models as possible that we develop are runnable on our website. This means that the code has to execute several hundred updates per cycle, and yet still run in the browser without lag. As we progress, we’re aiming to make things increasingly complex and explore as many different implementations of both the cell’s perception/interaction and the cell behaviour. These models will become increasingly taxing on the resources available, but from a HPC perspective, we’d like to push this as far as we can.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*2dl7Nq12z4sAE2xtAwLCaA.png" /><figcaption>Screenshot of Continuous Automata app</figcaption></figure><h3><strong>Reflections:</strong></h3><p><strong>What were some key lessons learned by the HPC team/or that you learnt from collaborating on this project?</strong></p><p>Keren (HPC) — In this project team, every individual has their own interest and knowledge geared toward a particular aspect of the project. So one of the first things new to me was learning how to work with a diverse range of knowledge bases and interests, from the machine learning component, web-development, mathematics component, GPU-programming, and the biological modelling. It’s been really interesting seeing how all these aspects work together, and figuring out how to adjust the workflow to best utilise the skills of each individual.</p><p><strong>Could you please provide three words to describe this project in retrospect?</strong></p><p>Nyan (AI) — Colourful, fast and challenging.</p><p>Colourful — If you see some of the demos that we have there’s a lot of really pretty patterns. A lot of worms quite frequently. I’ll say fast is another one — the HPC team is doing an incredible job making everything super fast. We give them a lot of code vomit, and they just make it work. And then lastly, I would say challenging. It’s very challenging at least on the AI side of things to train a neural network to replicate a rule set because it’s not just your standard supervised learning where you have a labelled image. You’re also kind of dealing with time series data in a sense so that makes training very difficult, and for this reason you get to explore out-of-the-box solutions.</p><p><em>Interested in learning more?</em></p><p><em>Check out these links:</em></p><p><a href="https://www.youtube.com/watch?v=ouipbDkwHWA&amp;t=125s"><em>Introduction to Conway’s Game of Life</em></a></p><p><a href="https://www.youtube.com/watch?v=xP5-iIeKXE8"><em>A video of Conway’s Game of Life, emulated in Conway’s Game of Life.</em></a></p><p><em>Lastly,</em></p><p><em>If you would like to access and play with our app, you can clone</em><a href="https://github.com/MonashDeepNeuron/Neural-Cellular-Automata.git"> <em>https://github.com/MonashDeepNeuron/Neural-Cellular-Automata.git</em></a>, go <em>to /CAs/Continuous and open life.html with live server (will need the live server extension on vscode).</em></p><p><strong>NCA Project Team Contributors:</strong></p><p>HPC — Alex Mai, Peter Cooper and Joshua Riantoputra</p><p>AI — Chloe Koe (Project Lead)</p><p><strong>Special thanks to our editors: </strong><em>Reya Jain and Regina Lu</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=e4e9a5f0e8a5" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Purveyors of Change: Confronting the New Challenges of Child Exploitation]]></title>
            <link>https://monashdeepneuron.medium.com/purveyors-of-change-confronting-the-new-challenges-of-child-exploitation-be744e4853fe?source=rss-19e1ddc97f32------2</link>
            <guid isPermaLink="false">https://medium.com/p/be744e4853fe</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Monash DeepNeuron]]></dc:creator>
            <pubDate>Mon, 18 Mar 2024 23:39:52 GMT</pubDate>
            <atom:updated>2024-03-19T01:54:40.820Z</atom:updated>
            <content:encoded><![CDATA[<p><strong><em>Parmeeta Siddique</em></strong></p><p>Monash DeepNeuron Law and Ethics Committee</p><p><strong>INTRODUCTION</strong></p><p>In recent years, we have seen the rapid growth of technology and at its centre, Artificial Intelligence (AI), and its innovative and <em>alarming </em>implementation in society. On September 5th 2023, Monash DeepNeuron’s Law and Ethics Committee collaborated with Monash University and AiLECS lab, who are at the forefront of developing AI so that it can assist law enforcement and be utilised to promote community safety, in the industry event, Protecting the Future: Combatting the Online Surge of Child Exploitation. This event seeks to leverage AI for the greater good and to raise awareness on the current failings of our justice system when it comes to protecting children from online crimes. However, despite AI’s untapped potential in the legal sector — it has also caused controversy due to various violations of data and privacy online.</p><p><strong>SETTING THE SCENE</strong></p><p>The event kicks off with a keynote from eSafety Commissioner Julie Inman Grant she explains how one of the biggest challenges we face with online activity is the lack of regulation and transparency. Online coercion has become an increasingly prevalent issue with sexual reports having tripled over the last year and 12% of reports to the eSafety Commission are self-produced Child Sexual Abuse Material (CSAM). Assistant Commissioner with the Australian Federal Police, Hilda Sirec offers insight to the landscape in which online perpetrators operate, and that CSA perpetrators are a notorious tech-savvy offender base which are always at the forefront of technological developments.</p><p>One of the most important takeaways from this event is the importance of proliferating the survivor voice and as a community, what we can do for victim-survivors of child exploitation and identifying the current issues which they face. From a legal perspective, our current systems of justice do not sufficiently protect victim-survivors in court. This is because victim-survivors experience secondary victimisation, which is where they are constantly forced to relive their trauma when navigating their case throughout the legal process. Furthermore, their cases may never be heard because of confusion when navigating the legal system such as determining the tangible evidence required to solidify their case, which jurisdiction should hear their case (Family, Civil or Criminal) and a general unwillingness to bring cases to law enforcement. This also brings up issues of privacy and identification, as victim-survivors and their CSAM become highly traded and often idolised by the CSA perpetrator online community, making it increasingly difficult to remove this explicit content on online platforms due to the dissemination on the clear web. Not only this, but there are several cultural factors which contribute to the diminishing of the survivor voice. There are many cases where the perpetrator is likely to be a close contact, for girls 88% of the time the perpetrator is someone they know, which often leads to cases of assault going unreported due to familial ties and relationship dynamics.</p><p>However, the most pressing concern is that there is an unwillingness by society at large, to elevate these voices. Rather than turn one’s head away from these atrocities, as a society, we must learn to confront and empathise with the struggles of victim-survivors because if we do not, then how long must they wait to receive their justice?</p><p><strong>STUDENT DEBATE</strong></p><p>The Law and Ethics Committee were largely involved in the debate portion of the event, featuring Elizabeth Perry, Leah Martinez, Merryn Cagney, and Parmeeta Siddique. The debate topic was whether the benefits of using Clearview AI and other similar technologies in law enforcement outweigh the risks.</p><p><em>Context</em></p><p>Clearview AI is a private corporation which uses facial recognition technology. The AI scrapes data across the web which is logged into a proprietary database containing over 30 billion facial images from public domains like social media, news outlets etc. to train itself. Then, when you upload a facial image, within seconds the AI will present you with links to that person’s images across the web.</p><p>However, Australian Information Commissioner and Privacy Commissioner Angelene Falk has found that Clearview AI, Inc. contravened Australians’ privacy as it was collecting biometric data without consent, breaching the Australian <em>Privacy Act 1988. </em>Clearview AI was then officially banned in Australia after Clearview AI Inc and Australian Information Commissioner [2023] AATA 1069 trial.</p><p><em>Conclusion of Debate</em></p><p>The affirmative argued largely on the basis that the current procedures that we have in place for law enforcement are insufficient. This is especially so, in relation to online child exploitation where Detective Inspector Jon Rouse from Queensland Police, said that victim identification is the most under-resourced skill in CSA. Using Clearview AI could greatly enhance and expedite victim identification and thus, discovering perpetrators exponentially, which allows law enforcement to use their resources elsewhere instead of an endless goose chase. Despite regulation around AI being under-developed, with proper oversight, it could become a valuable tool which promotes public security and safety.</p><p>To contend, the negative argued that the use of Clearview AI would fundamentally undermine our democratic rights as it completely disregards consent from users–and would remove online anonymity which is essential for living an authentic life. This technology could potentially be abused and would ultimately be a gross invasion to one’s entitlement of privacy. Furthermore, the improper use of this technology has led to false arrests and false positives due to prejudice in law enforcement and discrimination bias in the algorithm.</p><p>It was a strong debate, and although there was a general consensus from the audience in favour of using AI facial recognition technology, the negative reigned victorious in the debate.</p><p><strong>CONCLUSION</strong></p><p>When it comes to integrating AI into law enforcement, there is still much work to be done–especially when there is already a lot of distrust in its application from the general public. But even before that, there are social issues that are worth looking into and discussing further and making meaningful change. In 2021, the Australian Centre to Counter Child Exploitation received more than 33,000 reports of online child sexual exploitation–of which only 237 people were charged that year.</p><p>That is a 0.007% of turnovers. This is completely unacceptable.</p><p>We encourage you to read up further on CSA (we’ll have links below that you can read up on!) and if you are comfortable and want to help in the development of facial recognition technology for investigative purposes, you can upload your childhood photos onto My Pictures Matter (run by AiLECS labs) and help out today.</p><p>It is completely understandable to find these topics harrowing and difficult to face, but as a society, if we cannot face this reality, then these children will never be saved.</p><p><strong>ADDITIONAL RESOURCES</strong></p><ul><li>My Pictures Matter — <a href="https://mypicturesmatter.org/">https://mypicturesmatter.org/</a></li><li>DNA Doe Project — <a href="https://dnadoeproject.org/">https://dnadoeproject.org/</a></li><li>What is Child Sexual Abuse Material? — <a href="https://www.rainn.org/news/what-child-sexual-abuse-material-csam">https://www.rainn.org/news/what-child-sexual-abuse-material-csam</a></li><li>Secondary Victimisation — <a href="https://medium.com/@okonagata/beyond-the-crime-understanding-secondary-victimization-and-its-impact-on-victims-dd25e071a486">https://medium.com/@okonagata/beyond-the-crime-understanding-secondary-victimization-and-its-impact-on-victims-dd25e071a486</a></li><li>Increasing Trend of CSAM — <a href="https://www.aic.gov.au/sites/default/files/2021-02/ti616_production_and_distribution_of_child_sexual_abuse_material_by_parental_figures.pdf">https://www.aic.gov.au/sites/default/files/2021-02/ti616_production_and_distribution_of_child_sexual_abuse_material_by_parental_figures.pdf</a></li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=be744e4853fe" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[2023 APAC HPC-AI Competition]]></title>
            <link>https://monashdeepneuron.medium.com/2023-apac-hpc-ai-competition-fc8689acdfe3?source=rss-19e1ddc97f32------2</link>
            <guid isPermaLink="false">https://medium.com/p/fc8689acdfe3</guid>
            <dc:creator><![CDATA[Monash DeepNeuron]]></dc:creator>
            <pubDate>Wed, 13 Dec 2023 01:57:42 GMT</pubDate>
            <atom:updated>2023-12-13T02:05:17.497Z</atom:updated>
            <content:encoded><![CDATA[<p>In this competition, we are given two software, one HPC application, and one AI application. In our 2023 APAC competition, MPAS-Atmosphere was chosen as the HPC application, and Bloom as the AI application. Our task is to optimise these softwares to achieve as high a benchmark as possible while staying under compute resource constraints.</p><h3>The Team</h3><p>Monash DeepNeuron (MDN) is a Monash University undergraduate team that strives to empower, educate and engage students through hands-on, real world High-Performance Computing (HPC) and Artificial Intelligence projects.</p><p>As one of the projects this year, we’ve decided to participate in the APAC HPC-AI competition for a unique opportunity to be challenged, demonstrate the student’s skills, and learn from peers and experts from around the world.</p><p>This competition embodies our team’s commitment to innovation and excellence in the fields of HPC and AI. As we strive to create cutting-edge solutions and tackle complex challenges, participating in such a prestigious event will allow us to test the limits of our capabilities and push the boundaries of what is possible.</p><p>MDN opens a project interest form every semester, and our competition members were selected from a pool of students who expressed their interest in the APAC HPC-AI Competition.</p><p>Since we’ve entered the competition later than the official entry date, we’ve missed the online trainings, however the team were trained on the recorded training sessions, in addition to the 3-day NCI training session that ran on the Monash campus.</p><p>During the competition, we’ve held weekly hybrid meetings to catch up and sometimes work on the competition tasks during the meeting. The resources and communication were done on Microsoft Teams, and the code and documentation were pushed to GitHub. However, due to the nature of the competition tasks and the sizes of the resources, we’ve opted to create a shared directory in Gadi under our project, which made all the scripts and code resources available to the team members.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/724/1*7CJwmm5Rm_S5fq_RDJCidw.png" /><figcaption><strong><em>2023 APAC HPC-AI Team Members</em></strong></figcaption></figure><h3>Competition rules &amp; constraints</h3><p>The “BLOOM Inference” task focuses on accelerating the distributed inference of the BLOOM-176B model across two NCI Gadi V100x4 GPU compute nodes. Participants are expected to enhance inference speed without modifying the model itself, using recommendations like ds-inference shard-int8. Rules permit adjustments that don’t change the model but maintain output quality, and the competition assesses performance based on “Throughput per token, including tokenize” on designated GPU servers. Optimizations that reduce computation while retaining output are encouraged, but it’s crucial to ensure identical output. The competition limits AI application performance optimization scoring to specific projects (Ds, Ds-Zero, HF accelerate), but teams can explore other optimization solutions within certain guidelines.</p><p>MPAS-Atmosphere task required achieving the highest “simulation speed”, which are calculated by dividing the simulated time with the actual time it took for the computation. We had access to up to 32 Gadi CPU normal servers, and were not allowed to accelerate using GPUs. To ensure that the optimisation didn’t affect the “accuracy” of the simulation, we needed the values of the “global min, max w” and “global min, max u” lines in the output log to be identical across the runs.</p><h3>BLOOM</h3><h3>BLOOM Overview</h3><p>BLOOM(176B parameters) is a Large Language Model (LLM) designed for text generation. It’s trained on extensive text data using powerful computational resources and can produce coherent text in various languages and even programming languages, often indistinguishable from human writing. BLOOM can also handle text tasks beyond its training by framing them as text generation tasks. Its architecture closely resembles GPT-3(175B parameters), but with additional enhancements, as explained in the following sections.</p><h3>Libraries to Run BLOOM</h3><p>When deciding on a library to run and optimise the BLOOM language model, you should consider several key factors among the options of <a href="https://github.com/huggingface/transformers-bloom-inference/tree/main/bloom-inference-scripts">Accelerate, DeepSpeed, and DeepSpeed ZeRO</a>:</p><ul><li><strong>Parallelism Techniques: </strong>HF Accelerate uses Pipeline Parallelism, and DeepSpeed combines various parallelism techniques like tensor, pipeline, expert, and ZeRO-parallelism</li><li><strong>Ease of Use: </strong>Less loading time means experiment iteration can be quick and easier. Out of the above 3 models, DeepSpeed has the shortest loading time due to the sharding model</li><li><strong>Throughput and Scaling</strong>: DeepSpeed and DeepSpeed ZeRO offer advanced solutions for increasing throughput and scaling models across multiple GPUs and even nodes. DeepSpeed ZeRO’s sharding approach can significantly boost performance, potentially making it 8x to 16x faster when using 8 or 16 GPUs.</li><li><strong>Hardware Considerations:</strong> Best use of parallelism can be different depending on the hardware you have access to</li></ul><h3>Bloom strategy</h3><p><a href="https://github.com/huggingface/transformers-bloom-inference/tree/main/bloom-inference-scripts">Github link to baseline BLOOM inference</a></p><p>In our strategy, we aimed to optimize the performance of the BLOOM model using different parallelization techniques and resource configurations. Here’s a summary of the key steps and findings:</p><h3>Initial Assessment:</h3><p><a href="https://github.com/hpcac/2023-APAC-HPC-AI/blob/main/BLOOM-176B_Inference/BLOOM%20Inference%20PBS%20scripts%20-%20Build%20%26%20Run.md">We started by ensuring that the BLOOM 176B model</a> could run on the available resources. The baseline repository provided an int-8 quantization model of BLOOM, and model parallelization and KV cache settings were configured by default to scale according to the number of GPUs in the system.</p><h4>Exploring Parallelization:</h4><p>Once we confirmed that the model was functional, we ventured into optimizing parallelization. This involved adjusting parameters related to model parallelization and exploring alternative parallelism techniques such as ZeRO parallelism, pipeline parallelism, and speculative inference.</p><h4>Optimizing Batch Size:</h4><p>Considering the size of the BLOOM model and the compute resources at our disposal (4 GPUs on 2 nodes), we determined that the primary parameter to adjust was the <strong>batch size.</strong> We conducted an exhaustive search of all possible batch sizes and identified the maximum batch size, which was found to be 14 (12 for int8).</p><h4>Performance Evaluation:</h4><p>As a result of our efforts, we achieved a final throughput per token that was similar to the baseline BLOOM model in the repository.</p><h3>Speedup methods</h3><h4>Quantisation</h4><p>To optimize the resource-intensive requirements of running BLOOM-176B, a technique known as quantization is employed. Initially, this model necessitates 8 powerful GPUs to perform inference and an astounding 72 GPUs for fine-tuning. <a href="https://arxiv.org/abs/2208.07339">Through experimentation</a>, it’s been determined that instead of using the standard 4-byte FP32 precision, a nearly identical inference outcome can be achieved with 2-byte BF16/FP16 half-precision, effectively halving the model size. Further reduction is possible through 8-bit quantization, which employs quarter precision and reduces the model size to 1/4th of its original size. Quantization, however, involves rounding values from one data type to another, which can introduce some noise and result in information loss, resembling a form of lossy compression.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*9bxHqKFjrK41TeSn" /></figure><h3>KV Cache</h3><p>KV caching in the context of large language models (LLMs) involves the caching and reuse of intermediate data (Key and Value matrices in self-attention module) instead of recomputing them in every iteration of the decoding process. This caching approach reduces the computational complexity of the decoding mechanism, shifting the burden from computation to data transfers, which can lead to decreased arithmetic intensity.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*Xq_hx18yenr-OLLk" /></figure><h3>Speculative decoding</h3><p>Speculative decoding is an innovative approach within autoregressive language models, aimed at expediting text generation while maintaining the same output from the large model. This method involves employing a smaller approximation model to generate token sequences, which are then scrutinised by a larger model. If the larger model deems the proposed tokens unnatural, it can reject the smaller model’s suggestions. This technique mimics direct sampling from the larger model, offering the advantage of faster evaluation of token sequence likelihood. It is particularly beneficial for sequences where generating some tokens is easier than others. Speculative decoding can be triggered when signals, like argmax logit magnitude, indicate the smaller model’s output might be inadequate, avoiding unnecessary speculation. However, it can be challenging to scale the code for multiple batches and integrate it with other speed-up techniques. It’s also crucial to experiment with various model sizes to determine the best fit for a target model with 176 billion parameters. Approximation models can be selected from pretrained models like bloom-560m, bloom-1b1, bloom-1b7, bloom-3b, and bloom-7b1. The approximation model’s size should be approximately ten times smaller than the target model <a href="https://arxiv.org/abs/2211.17192">according to the paper</a>, although it may not be entirely compatible with existing Bloom models.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/975/0*ryNo07HLdiB_E65H" /></figure><h3>Parallel Inference of BLOOM using Multi GPU</h3><h4>Batch size</h4><p>When using int8 data type for inference, float16 and int8 models show no significant difference. Increasing the batch size by a factor of 2 reduces throughput per token by the same factor, although it doesn’t change the total tokenization and token generation; instead, it enhances throughput per token by processing more tokens in one inference. In deep learning, the batch size determines how many data samples are processed in a single forward and backward pass, potentially improving GPU utilization but possibly demanding more memory. It’s important to note that batch size should not be confused with data parallelism, a technique where multiple copies of the same model operate on distinct data subsets.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*mva7rza3JgKYdWH6" /></figure><h3>Model(Tensor) Parallelism</h3><p>Tensor Parallelism (TP) is a technique employed in large-scale language model training, where each GPU processes a section of a tensor and combines the complete tensor only when necessary for operations that require it. By splitting weight matrices and parallelizing matrix multiplications, TP optimizes parallel processing in deep learning models, such as transformers, enabling more efficient use of GPU resources. It is particularly valuable for models with numerous parameters and can be further enhanced with sequence parallelism for specific operations. However, it does require fast interconnects between GPUs and may be constrained by the number of GPUs on a node.</p><p>In the transformer model, each layer comprises self-attention and MLP, which can be concurrently processed using tensor parallelism. The tensor parallel size needs to be divisible by the number of heads in the transformer model for DeepSpeed. Since BLOOM has 112 heads and we have 8 GPUs at our disposal, the available tensor parallel parameters are 1, 2, 4, and 8. However, due to the model’s size, BLOOM cannot fit into fewer than 8 GPUs, making 8 the determined tensor parallel parameter.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*wL4LsJvHlXhtwyDC" /></figure><h3>Other Parallelism Methods</h3><h4>ZeRO Data Parallelism</h4><p>Please note that default data parallelism cannot be applied to BLOOM as data parallelism assumes one model can fit into 1 GPU and run multiple streams of models on multiple distributed data simultaneously. In our case, Bloom barely fits 8 GPUs by separating them and will not fit into 1 GPU.</p><p>ZeRO Data Parallelism (ZeRO-DP) is a straightforward approach to distributed deep learning. In ZeRO-DP, traditional Data Parallelism (DDP) is used, but with an important modification. Instead of each GPU replicating the entire model parameters, gradients, and optimizer states, each GPU only holds a portion or slice of these components. When a specific layer’s parameters are required during runtime, all GPUs synchronize to exchange the missing parts, ensuring that each GPU has access to the complete layer parameters. This method optimizes memory usage and enables efficient parallel processing in distributed training.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*v_Nd-cejmCyRTaBh" /></figure><h4>Limitation</h4><ul><li><em>Long loading time- </em>We had to use big science model 176B parameters instead, which takes at least 40 mins to load.</li><li><em>Memory error- </em>V100s with 32GB memory was unable to load Big Science model without cpu offloading</li><li>CUDA memory error</li></ul><p>The testing that we did was only a smaller 560m parameter model which takes less memory and less loading time. Even though DeepSpeedZero was able to process more batch size, DeepSpeed had lower(more efficient) throughout per token.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*KFlDAsmNsntu-MUQ" /></figure><h3>Pipeline Parallelism</h3><p>Pipeline Parallelism, involves distributing groups of model layers across multiple GPUs and passing data between them as if they form a single composite GPU. This approach vertically slices the layers of a model, where each set of layers resides on a different GPU. During forward and backward passes, data is transferred between these GPUs as needed. However, the limitations involve that most GPUs are idle at any given time, making it less efficient and potentially leading to overhead when copying data between devices. While it can potentially increase available memory, it may not be as efficient as using larger single GPUs for some tasks, and it can introduce communication overhead, especially when GPUs are located on different compute nodes.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*5VrsVoK1kOwLmFye" /></figure><h3>MPAS</h3><h4>Overview</h4><p>The Model for Prediction Across Scales (MPAS), which is a collaborative project for developing atmosphere, ocean and other earth-system simulation components for use in climate, regional climate and weather studies.</p><p>MPAS uses unstructured Voronoi meshes and C-grid discretization, used as the basis for higher-resolution climate and mesoscale atmosphere simulations.</p><p>We’ve used the same versions of the libraries and software as used in the <a href="https://github.com/hpcac/2023-APAC-HPC-AI/blob/main/MPAS-Atmosphere_benchmark_10km/MPAS-Atmosphere%20PBS%20scripts%20-%20Build%20%26%20Run.md">provided script</a>:</p><ul><li>HDF5: 1.12.1</li><li>PNETCDF: 1.12.3</li><li>NETCDF: 4.8.1</li><li>PIO: 2.5.9</li><li>NETCDF_FORTRAN: 4.6.1</li><li>METIS: 5.1.0</li><li>MPAS-Model-7.3</li><li>Metis: 5.1.0</li></ul><h3>Strategy</h3><p>There were three main parts of MPAS-Atmosphere we can potentially modify to improve the performance:</p><ul><li>MPAS-Model source code</li><li>Namelist.atmosphere (strictly the PIO configuration as otherwise it would modify the model/physics, invalidating our results)</li><li>Compile optimisations</li></ul><p>Only the modification of the PIO configuration in the namelist file and the compilation optimisation has worked successfully.</p><p><strong>For the PIO configuration, the default was:</strong></p><p><em>config_pio_num_iotasks = 0</em></p><p><em>config_pio_stride = 1</em></p><p><em>The tweaked parameters were:</em></p><p><em>config_pio_num_iotasks = 48</em></p><p><em>config_pio_stride = 8/16/32</em></p><p>What this tweak allowed for was an increase in the effective process read rate in the IO Time by 18 MB/s, presumably due to the increased stride allowing for more effective/efficient transmission of data.</p><p>While doing the testing, the 48 core per node parameter tested the best.</p><p>When we modified the stride from 1 to 8, it managed to reduce the I/O overhead.</p><p><strong>Before:</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/829/0*TSQcXTZ9cJcl24mI" /></figure><p><strong>After:</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/655/0*HD8eaF-qd5g5TVJ6" /></figure><p>C/C++ and Fortran compilers offers optimisation flags denoted by `-O`. By default, this is set to O3, which is the highest optimisation which follows the standards compliance. However, we can also use `-Ofast` which disregards the standards compliance and enables further optimisation which turns on <a href="https://gcc.gnu.org/onlinedocs/gcc/Optimize-Options.html">-ffast-math, -fallow-store-data-races and the Fortran-specific -fstack-arrays</a>.</p><p>Since our Gadi normal CPU instances are using the Intel Platinum 8268 processor, which is part of the Caskade Lake family, these processors have AV512 instructions sets available. We’ve used the -mavx512f flag in our compilation to enable this feature too.</p><p><strong>The Default Compilation Flags:</strong></p><p><em>FFLAGS_OPT = -O3 -ffree-line-length-none -fconvert=big-endian -ffree-form</em></p><p><em>CFLAGS_OPT = -O3</em></p><p><em>CXXFLAGS_OPT = -O3</em></p><p><em>LDFLAGS = -O3</em></p><p><em>The Tweaked Configuration:</em></p><p><em>FFLAGS_OPT = -Ofast -ffree-line-length-none -fconvert=big-endian -ffree-form -mavx512f</em></p><p><em>CFLAGS_OPT = -Ofast -mavx512f</em></p><p><em>CXXFLAGS_OPT = -Ofast -mvx512f</em></p><p><em>LDFLAGS = -Ofast -mavx512f</em></p><p>All these tweaks tweak decreases the time integration by 18% in 8 node, when compared to the benchmark, increasing the total simulation speed. However, the improvement did not scale well, having improvement of around 5% in 32 node instance.</p><h3>Results</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/342/1*bNaYk45LccFF9Lu2JfAbDA.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/328/1*BbXhtU8SEfZ6CvdgpPxLew.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/380/1*In89oBAOGjQb9HiGOlV5iQ.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/389/1*bQxDwcBSMcQZ9kk6-s3nIA.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*yZ_biWxzk2Laj_oog1XkKg.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/609/1*OKZF4suGjoYIIHdhMr80bQ.png" /></figure><p>We have also attempted the following, however with no success:</p><ul><li>Compiling libraries and MPAS-Model with IntelMPI instead of OpenMPI -&gt; compilation error</li><li>Compiling MPAS-Model with OpenMP -&gt; No performance improvement/worst performance</li><li>Algorithmic/code efficiency improvements to the MPAS-Model (by directly editing the Fortran source code) -&gt; either compilation error, or producing numerical errors</li><li>Profiling individual functions within the MPAS-Model and MPI calls -&gt; unable to retrieve satisfactory results/error during runtime</li><li>Adding –map-by core –bind-by core flags in the PBS script that runs the simulation: No change in performance</li></ul><h3>Conclusion</h3><p>The 2023 APAC HPC-AI Competition was a challenging yet rewarding experience for the MDN/Monash team. Despite entering the competition later than the official entry date, the team exhibited a remarkable dedication to learning and applying optimisation techniques for the MPAS-Atmosphere and Bloom. However, there are lots of rooms for improvements for both applications.</p><p>While the team encountered some technical challenges, such as memory limitations and long loading times, they continuously strived to find innovative solutions and demonstrated resilience in the face of adversity. However, for MPAS in particular, we have encountered many issues with compiling the software with some of the optimisation techniques, combined with the academic commitments, we were able to spend satisfactory amount of time to debug and explore further optimisations. We also saw some deficiencies in the Fortran and mathematical background that was used in the MPAS, which further slowed down our optimisation efforts.</p><p>The competition provided an exceptional platform for the team to familiarise themselves with the intricacies of various cutting-edge optimisation methodologies. The team delved into deep learning optimisation techniques that were particularly joyful to learn about.</p><p>Looking ahead, the MDN team is committed to incorporating the lessons learned from this competition into their future projects and competitions. We aim to refine our technical capabilities, enhance collaborative skills, and explore additional avenues for optimising software performance under resource constraints.</p><p>The team extends its sincere gratitude to the organisers for providing this invaluable opportunity to participate in an international contest of this stature. The competition not only challenged their technical prowess but also fostered a spirit of camaraderie and innovation among the team members, leaving a lasting impact on their academic and professional journey.</p><p><em>Authored by Yuki Kume, Caleb Ooi, Aditya Desai, Allister Lim, Yusuke Miyashita</em></p><h3>Reference</h3><p>1. Hugging Face. (2022). The Technology Behind BLOOM Training. Retrieved from <a href="https://huggingface.co/blog/bloom-megatron-deepspeed">https://huggingface.co/blog/bloom-megatron-deepspeed</a></p><p>2. Hugging Face. (2022). A Gentle Introduction to 8-bit Matrix Multiplication for transformers at scale using Hugging Face Transformers, Accelerate and bitsandbytes. Retrieved from <a href="https://huggingface.co/blog/hf-bitsandbytes-integration">https://huggingface.co/blog/hf-bitsandbytes-integration</a></p><p>3. Tim D. (2022). LLM.int8(): 8-bit Matrix Multiplication for Transformers at Scale. arXiv preprint arXiv:2208.07339. Retrieved from <a href="https://arxiv.org/abs/2208.07339">https://arxiv.org/abs/2208.07339</a></p><p>4. Axelera AI. (2023). Decoding Transformers on Edge Devices. Retrieved from <a href="https://www.axelera.ai/decoding-transformers-on-edge-devices/">https://www.axelera.ai/decoding-transformers-on-edge-devices/</a></p><p>5. Yaniv L. (2022). Fast Inference from Transformers via Speculative Decoding. arXiv preprint arXiv:2211.17192. Retrieved from <a href="https://arxiv.org/abs/2211.17192">https://arxiv.org/abs/2211.17192</a></p><p>6. Hugging Face. (2022). bloom-inference-scripts. Retrieved from <a href="https://github.com/huggingface/transformers-bloom-inference/tree/main/bloom-inference-scripts">https://github.com/huggingface/transformers-bloom-inference/tree/main/bloom-inference-scripts</a></p><p>7. Microsoft Research. (2023). ZeRO &amp; DeepSpeed: New system optimizations enable training models with over 100 billion parameters. Retrieved from <a href="https://www.microsoft.com/en-us/research/blog/zero-deepspeed-new-system-optimizations-enable-training-models-with-over-100-billion-parameters/">https://www.microsoft.com/en-us/research/blog/zero-deepspeed-new-system-optimizations-enable-training-models-with-over-100-billion-parameters/</a></p><p>8. Weng, L. (2021). How to Train Really Large Models on Many GPUs? Retrieved from <a href="https://lilianweng.github.io/posts/2021-09-25-train-large/">https://lilianweng.github.io/posts/2021-09-25-train-large/</a></p><p>Colossal AI. (2023.). Paradigms of Parallelism. Retrieved from <a href="https://colossalai.org/docs/concepts/paradigms_of_parallelism/">https://colossalai.org/docs/concepts/paradigms_of_parallelism/</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=fc8689acdfe3" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Mental Health Chatbots]]></title>
            <link>https://monashdeepneuron.medium.com/mental-health-chatbots-1581b845e3b3?source=rss-19e1ddc97f32------2</link>
            <guid isPermaLink="false">https://medium.com/p/1581b845e3b3</guid>
            <dc:creator><![CDATA[Monash DeepNeuron]]></dc:creator>
            <pubDate>Sun, 22 Oct 2023 04:29:16 GMT</pubDate>
            <atom:updated>2023-10-22T04:29:16.344Z</atom:updated>
            <content:encoded><![CDATA[<h3>Mental Health Chatbots</h3><p>Although Australia has begun to settle into its ‘post COVID-19 era’, levels of psychological distress and life satisfaction amongst Australians are still yet to recover to what they were pre-pandemic. <a href="https://www.deepneuron.org/lae-blogs/2023/10/20/mental-health-chatbots#links">1</a> Some have referred to this as the ‘shadow mental health pandemic’. <a href="https://www.deepneuron.org/lae-blogs/2023/10/20/mental-health-chatbots#links">2</a></p><p>Unfortunately, Australia is simultaneously in the midst of a ‘mental health crisis’ as the mental healthcare sector struggles to cope with the increased demand for services. <a href="https://www.deepneuron.org/lae-blogs/2023/10/20/mental-health-chatbots#links">3</a> A 2023 survey of 1,190 Australian Psychological Society members nationwide found that 77% of respondents reported increased or consistent wait times since February 2022, <a href="https://www.deepneuron.org/lae-blogs/2023/10/20/mental-health-chatbots#links">4</a> at which time the average wait time to see a psychologist was between three to six months. <a href="https://www.deepneuron.org/lae-blogs/2023/10/20/mental-health-chatbots#links">5</a></p><p>In the fight against the so-called shadow mental health pandemic, artificial intelligence (‘AI’) chatbots have emerged as a potential means of supplementing mental healthcare services to achieve better patient outcomes. <a href="https://www.deepneuron.org/lae-blogs/2023/10/20/mental-health-chatbots#links">6</a></p><p>These AI chatbots are conversational agents that use natural language processing (‘NLP’) to simulate human-like conversation with users. <a href="https://www.deepneuron.org/lae-blogs/2023/10/20/mental-health-chatbots#links">7</a> This allows them to, amongst other things, deliver mental healthcare interventions such as those based on cognitive behavioural therapy (‘CBT’) principles. <a href="https://www.deepneuron.org/lae-blogs/2023/10/20/mental-health-chatbots#links">8</a></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/600/0*u8xJHjxGLsy3DTI4.jpg" /></figure><p>[Wysa is one of the leading apps in this field: <a href="https://www.verdict.co.uk/mental-health-chatbots/">https://www.verdict.co.uk/mental-health-chatbots/</a>]</p><p><strong>The Benefits</strong></p><p>It has been argued that AI chatbots will improve equity and social inclusion in mental healthcare by contributing to the affordability, efficiency and convenience of receiving support.<a href="https://www.deepneuron.org/lae-blogs/2023/10/20/mental-health-chatbots#links">9</a></p><p>For those unable or unwilling to access traditional mental healthcare, there have been promising early findings of both acceptance towards AI chatbots as well as positive health outcomes.<a href="https://www.deepneuron.org/lae-blogs/2023/10/20/mental-health-chatbots#links">10</a> Users have provided feedback that AI chatbots, in comparison with mental healthcare professionals, provide an unbiased, judgement-free zone that can be accessed 24/7 and for durations that are as little or as long as required. Further, they have been found to decrease depressive symptoms, support the development of interpersonal skills, reduce stress, teach self-compassion, and alleviate feelings of loneliness.<a href="https://www.deepneuron.org/lae-blogs/2023/10/20/mental-health-chatbots#links">11</a></p><p>The developers of these chatbots reiterate that it is not their intention to replace human clinicians, but to supplement their services — especially where there is insufficient supply to meet demand for support (such as in Australia).</p><p><strong>The Risks</strong></p><p>However, some users as well as experts have voiced their hesitancies with using AI chatbots for matters of such a sensitive and personal nature as mental healthcare.</p><p>Unlike a human clinician, an AI chatbot is not able to perceive live social cues, body language or evidence of behavioural patterns.<a href="https://www.deepneuron.org/lae-blogs/2023/10/20/mental-health-chatbots#links">12</a> For example, if a user were to tell an AI chatbot that they were ‘sad’, the chatbot would not be able to adequately assess the user’s level of sadness or the way in which these feelings impact the user’s daily functioning unless this information was volunteered.<a href="https://www.deepneuron.org/lae-blogs/2023/10/20/mental-health-chatbots#links">13</a> On the other hand, a human clinician might be able to observe that the patient appears to not have showered in several days, or that they are unable to make eye contact and are slouching, which indicates a level of impairment in functioning.<a href="https://www.deepneuron.org/lae-blogs/2023/10/20/mental-health-chatbots#links">14</a></p><p>Problematically, various AI chatbots have been found to produce inappropriate and potentially very damaging outputs in response to users seeking their support. For example, America’s National Eating Disorder Association (‘NEDA’) sought to replace its helpline counsellors with an AI chatbot referred to as ‘Tessa’.<a href="https://www.deepneuron.org/lae-blogs/2023/10/20/mental-health-chatbots#links">15</a> Unfortunately, earlier this year, it was forced to take Tessa down as the AI chatbot had begun giving weight loss tips to users who were seeking support for eating disorders.<a href="https://www.deepneuron.org/lae-blogs/2023/10/20/mental-health-chatbots#links">16</a></p><p>It is well recognised that AI chatbots are severely lacking in their ability to recognise when a user is in crisis.<a href="https://www.deepneuron.org/lae-blogs/2023/10/20/mental-health-chatbots#links">17</a> In addition to providing inadequate crisis support, there is also the adjacent issue of the social isolation that may arise from a user’s over-reliance on these technologies immediately prior to the onset of a crisis.<a href="https://www.deepneuron.org/lae-blogs/2023/10/20/mental-health-chatbots#links">18</a> It has been found that some users enjoyed the ‘relationship’ they had with the AI chatbots so much, to the extent that they believed that they could replace their friends and family members.<a href="https://www.deepneuron.org/lae-blogs/2023/10/20/mental-health-chatbots#links">19</a></p><p><strong>Balancing the Benefits and Risks</strong></p><p>Although AI chatbots for use in the mental healthcare sector were arguably developed with the best of intentions to supplement a struggling workforce (and are undoubtedly of utility in this regard), their implementation and use must be closely overseen. Unfortunately, at this stage, regulations specific to the use of AI chatbots in the mental healthcare sector are yet to be developed.</p><p>The European Parliament’s draft ‘Artificial Intelligence Act’ (‘draft AI Act’), which seeks to establish a regulatory framework for the use of AI in the European Union, will hopefully provide some much needed guidance on how the responsible oversight of AI may also be achieved in Australia in the future. It is thought that for the purposes of the draft AI Act, AI chatbots will likely fall within the scope of medical devices, in which case they will be classified as ‘high risk’ systems.<a href="https://www.deepneuron.org/lae-blogs/2023/10/20/mental-health-chatbots#links">20</a> This will require such technologies to, amongst other things, have a risk management system in place.<a href="https://www.deepneuron.org/lae-blogs/2023/10/20/mental-health-chatbots#links">21</a></p><p>In the meantime, the mental healthcare sector’s responsible employment of these AI chatbots will be critical. It must be remembered that these technologies are not intended to replace human clinicians or existing crisis service delivery methods (such as NEDA’s helpline counsellors), but to supplement their services.</p><p><em>Authored by Johanna Frauenberger, Monash DeepNeuron Law and Ethics Committee</em></p><p>[1] <a href="https://whataustraliathinks.org.au/data_story/reflecting-on-the-pandemic-how-has-australias-mental-health-faired/">Reflecting on the pandemic — How has Australia’s mental health faired</a></p><p>[2] <a href="https://www.sciencedirect.com/science/article/pii/S2666560321000177">AI chatbots cannot replace human interactions in the pursuit of more inclusive mental healthcare</a></p><p>[3] <a href="https://psychology.org.au/about-us/news-and-media/aps-in-the-media/2023/analysing-australias-mental-health-crisis">Analysing Australia’s mental health crisis</a></p><p>[4] <a href="https://psychology.org.au/about-us/news-and-media/media-releases/2023/national-survey-sounds-alarm-for-patients-and-psyc">National survey sounds alarm for patients and psychologists after mental health funding cuts</a></p><p>[5] <a href="https://psychology.org.au/for-members/news-and-updates/news/2022/australians-need-psychological-help-more-than-ever">Australians need psychological help more than ever</a></p><p>[6] Ibid.</p><p>[7] <a href="https://ieeexplore.ieee.org/abstract/document/9801962">The Role of AI Chatbots in Mental Health Related Public Services in a (Post) Pandemic World: A Review and Future Research Agenda</a></p><p>[8] <a href="https://www.sciencedirect.com/science/article/pii/S2214782922000021">Using AI chatbots to provide self-help depression interventions for university students: A randomized trial of effectiveness</a></p><p>[9] <a href="https://www.sciencedirect.com/science/article/pii/S2666560321000177">AI chatbots cannot replace human interactions in the pursuit of more inclusive mental healthcare</a></p><p>[10] <a href="https://www.ncbi.nlm.nih.gov/pmc/articles/PMC10242473/">An Overview of Chatbot-Based Mobile Mental Health Apps: Insights From App Description and User Reviews</a></p><p>[11] Ibid.</p><p>[12] Ibid.</p><p>[13] Ibid.</p><p>[14] Ibid.</p><p>[15] <a href="https://ifstudies.org/blog/the-ramifications-of-ai-for-mental-health">The Ramifications of AI for Mental Health</a></p><p>[16] Ibid.</p><p>[17] Ibid.</p><p>[18] Ibid.</p><p>[19] Ibid.</p><p>[20] <a href="https://link.springer.com/chapter/10.1007/978-3-031-18697-4_21">Legal Issues of the Use of Chatbot Apps for Mental Health Support</a></p><p>[21] Ibid.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1581b845e3b3" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Cultural Impact of AI in Art]]></title>
            <link>https://monashdeepneuron.medium.com/the-cultural-impact-of-ai-in-art-a4b6e745a7b3?source=rss-19e1ddc97f32------2</link>
            <guid isPermaLink="false">https://medium.com/p/a4b6e745a7b3</guid>
            <category><![CDATA[culture]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[art]]></category>
            <category><![CDATA[technology]]></category>
            <dc:creator><![CDATA[Monash DeepNeuron]]></dc:creator>
            <pubDate>Sat, 19 Aug 2023 06:37:56 GMT</pubDate>
            <atom:updated>2023-08-19T06:37:56.470Z</atom:updated>
            <content:encoded><![CDATA[<p><strong><em>Parmeeta Siddique</em></strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/686/0*UIJI6nU-PHh8-PAh.png" /></figure><p><strong><em>[Above image shared by Twitter user Whyenn (@WhyTheEnn)]</em></strong></p><p>Artificial Intelligence (AI) began stirring controversy in 2022 with the release of Midjourney, which utilises AI and text-to-image generation to allow millions of Discord users to generate art. Although you might not see six-fingered ladies with such high frequency anymore, AI art has — without a doubt — opened up the possibilities of AI in the creative industry with its ability to produce music, art, poetry and even anime (Corridor Digital’s ‘Rock, Paper, Scissors’ for the brave).</p><p>The debate amongst commentators revolves around the role that AI plays in art and artists, and how it affects the art community at large in a capitalist-consumerist society. Does AI art undermine the sanctity of art and subsequently artists themselves? Or does it have the capacity to re-define creativity as we know it?</p><p><strong><em>How does ‘AI Generated’ Art Work?</em></strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/554/0*HZaUM-QUgZq7o2yY.png" /></figure><p><strong><em>[Prompt: “Two pigs staring up at the moon, watercolour painting” (DALL-E 2)]</em></strong></p><p>‘Generated AI art’ comes from generative AI programs which can produce images, such as photographs, paintings or drawings that look as though they have been created by humans. Generative AI programs such as Midjourney, DALL-E, Stable Diffusion and others, use diffusion models, where programmers feed the algorithm real images (often without the consent of the owners of those said images) and the neural network learns the common features and patterns associated with said images, and tries to implement them into the end product. This step is repeated over and over until nothing is left to be seen of the original image, generating ‘something new’ but still fits the prompt. The goal of these models is to learn from the input of data, so that it can distinguish what it means to generate, for example, a dog as opposed to a cat. Once it can navigate and refine the initial random distribution of data, it eventually removes the noise from data, and creates a realistic sample of what was prompted, this process of denoising is what is known as diffusion.</p><p>It’s not a Frankenstein creation that combines segments of art (like a nose from Picasso combined with an eye from Leonardo Da Vinci), but a blend of stylistic characteristics to form one image <a href="https://www.deepneuron.org/lae-blogs/2023/3/31/chatgpt-in-education-jta5h#_ftn1">[1]</a>. This is why it looks like these images have been created with meaning and purpose, despite an AI not knowing what that might entail.</p><p><strong><em>A Brief History of AI Art</em></strong></p><p>AI art existed long before its rise in the 2020’s, where now it lets you become a cute anime girl using TikTok’s new AI filter (its uses certainly have changed…)</p><ul><li>1960s — Harold Cohen (a British artist specialising in art and computing) developed algorithms allowing computers to imitate drawing by hand–calling the project AARON. Although it was intended to paint specific objects, the machine generated pictures that he had never anticipated before.</li><li>2000s — A field that was once exclusive to those only with high technical backgrounds, became more open to the public due to the rapid advancements of this era. This led to the development of programs such as ImageNet and Google DeepDream, creating an intersection between art and code where computer software can recognise aesthetic patterns and visual objects.</li><li>2014 — Generative Adversarial Networks (GANs) were introduced by Ian Goodfellow (American computer scientist and engineer), involving two neural networks: the ‘generator’ and the ‘discriminator’ which informs if the input received is real or from the ‘generator’ network. These models are trained until the discriminator can be fooled enough times — where it becomes unique enough.</li></ul><p>This complex history goes to show that while art has always been fascinated by futuristic cybernetics, the technological spectacles and the endless possibilities it offers… The same can be applied in the reverse, as technology has also been intrigued with entering the supposedly elusive realm, that is art. Both art and technology are driven by people, and yet art is always seen as inexplicably more ‘human’ than technology is. The intersection and exploration of these two factions, can allow for evolution in their respective fields, so why is technology creating art seen as an invasion?</p><p><strong><em>What is ‘Art?’</em></strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/733/0*ow2splme5Hm2mabL.png" /></figure><p><strong><em>[AI generated art, the Théâtre D’opéra Spatial by Jason M. Allen (Midjourney)]</em></strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/324/0*UtgTRfXCs7QG2Fd6.png" /></figure><p><strong><em>[‘Comedian’ by Maurizio Cattelan]</em></strong></p><p>If art is a form of expression that includes the absurd — is ‘art’ created by an AI, really so offensive?</p><p>One of the points in this debate is the arguable lack of skill that comes with generated AI art. Especially in relation to transparency. Théâtre D’opéra Spatial by Jason M. Allen, created using Midjourney, won first place in the ‘digital arts/digitally-manipulated photography’ category at the Colorado State Fair Fine Arts Competition. It took over 80 hours to complete with over 900 iterations generated <a href="https://www.deepneuron.org/lae-blogs/2023/3/31/chatgpt-in-education-jta5h#_ftn1">[2]</a></p><p>This suggests that although it is easy to create something acceptably aesthetic, more curation and decision making is required throughout the process. However, this does not outweigh the time artists put into developing practical and physical skill, with the inclusion of decision making, in addition to all the materials needed, to create their art.</p><p>The biggest issue with AI art is how it trains itself from other people’s works without consent–and although it will not create an identical image, it can mimic that person’s style. However, art style cannot be copyrighted–that’s why artists are upset, their works are not well protected by the law.</p><p>However, if the machine alters an image so much that it cannot be directly identified with a single piece of art, how is that different from people being inspired from each other every day, and implementing those ideas in their own art? A survey conducted by Yale University tested its students on whether or not it could differentiate AI generated images and human-made art. Out of the 504 undergraduates, people could correctly identify the difference 54% of the time <a href="https://www.deepneuron.org/lae-blogs/2023/3/31/chatgpt-in-education-jta5h#_ftn1">[3]</a> .</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/363/0*yFhhk1_o0Yo9S7RO.png" /></figure><p><strong><em>[prompt: “an empty teacup, oil painting” (DALL-E 2)]</em></strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/486/0*JXh1tX9eF2lOHG_9.png" /></figure><p><strong><em>[Artist Ken Lee’s long exposure night photo]</em></strong></p><p>So, is AI art, really art? Probably. Despite not being created by humans, it can mimic human emotions, techniques, skills, imagination and more, because it’s been trained from human data.</p><p>What’s important though, is the uneasy feeling that comes with this fact. Should AI art be seen in the same category as human-made art? No. Not in a capitalist society.</p><p><strong><em>Capitalism and Art</em></strong></p><p>Yes, the juicy controversy of technology invading the art world is worthy of exploration, but at the end of the day: we live in a society.</p><p>Referring to Australia’s ethical principles of AI <a href="https://www.deepneuron.org/lae-blogs/2023/3/31/chatgpt-in-education-jta5h#_ftn1">[4]</a> , AI art has already caused a multitude of issues. Issues of transparency and explainability–AI programs training themselves on art without the artists’ consent. Issues of fairness–most of these AI programs have been developed by Western companies and hence, have eurocentric concerns. MIT Media Lab researcher Joy Buolamwini identified how most AI facial-recognition software cannot recognise dark-skinned faces or certain ethnic features which could lead to further discrimination <a href="https://www.deepneuron.org/lae-blogs/2023/3/31/chatgpt-in-education-jta5h#_ftn1">[5]</a>. Not to mention the invasion of human autonomy concerning Deepfakes.</p><p>AI art impacts our human-centred values and societal wellbeing, and it’s important to interrogate why.</p><p>AI art is comparable to the industrial revolution, technology replacing manual labour, which supposedly improves our quality of life. However, this analogy cannot be applied to art which uses creative labour. Artists, illustrators, animators, are already devalued in capitalist systems as it exists in a liminal space between leisure and labour <a href="https://www.deepneuron.org/lae-blogs/2023/3/31/chatgpt-in-education-jta5h#_ftn1">[6]</a> . The introduction of AI art will likely replace commercial art in a fast-paced consumerist society <a href="https://www.deepneuron.org/lae-blogs/2023/3/31/chatgpt-in-education-jta5h#_ftn1">[7]</a> . Instant gratification, mass production, and abundant consumption, these are factors that will ultimately affect how human-made art will be devalued, and justify paying them even less than before.</p><p>Of course we can consider how AI might help artists, but can we really rely on conglomerates and corporations, not taking the quicker, cheaper, and easier route?</p><p><strong><em>Conclusion</em></strong></p><p>Amongst gangly limbs and promiscuous body proportions, one cannot deny the capacity AI has in the creative industry. Although it’s hilarious to see Kanye West and Drake performing a soulful duet of ‘Breaking Free’ on TikTok, the fact that AI has increasingly become more present in the creative industry, whether it comes to voice acting, songs, writing, art and more, could be something of concern. Creative labour is valuable but vulnerable in a capitalist society, and it’s up to us to decide how much we value it.</p><p>Art has a story to tell. A person who may or may not be someone like you, has carefully translated their livelihood and experiences into making the fundamentals of that story, and every detail they include, may have meaning that will resonate with you. But an AI will never owe you such an explanation.</p><p><strong><em>Authored by</em> <em>Parmeeta Siddique, Monash DeepNeuron Law and Ethics Committee</em></strong></p><p><a href="https://www.deepneuron.org/lae-blogs/2023/3/31/chatgpt-in-education-jta5h#_ftnref1">[1]</a> <a href="https://www.theguardian.com/technology/2022/nov/12/when-ai-can-make-art-what-does-it-mean-for-creativity-dall-e-midjourney">https://www.theguardian.com/technology/2022/nov/12/when-ai-can-make-art-what-does-it-mean-for-creativity-dall-e-midjourney</a></p><p><a href="https://www.deepneuron.org/lae-blogs/2023/3/31/chatgpt-in-education-jta5h#_ftnref1">[2]</a> <a href="https://edition.cnn.com/2022/09/03/tech/ai-art-fair-winner-controversy/index.html">https://edition.cnn.com/2022/09/03/tech/ai-art-fair-winner-controversy/index.html</a></p><p><a href="https://www.deepneuron.org/lae-blogs/2023/3/31/chatgpt-in-education-jta5h#_ftnref1">[3]</a> <a href="https://yaledailynews.com/blog/2023/01/23/what-ai-art-means-for-society-according-to-yale-experts/">https://yaledailynews.com/blog/2023/01/23/what-ai-art-means-for-society-according-to-yale-experts/</a></p><p><a href="https://www.deepneuron.org/lae-blogs/2023/3/31/chatgpt-in-education-jta5h#_ftnref1">[4]</a> <a href="https://www.industry.gov.au/publications/australias-artificial-intelligence-ethics-framework/australias-ai-ethics-principles">https://www.industry.gov.au/publications/australias-artificial-intelligence-ethics-framework/australias-ai-ethics-principles</a></p><p><a href="https://www.deepneuron.org/lae-blogs/2023/3/31/chatgpt-in-education-jta5h#_ftnref1">[5]</a> <a href="https://medium.com/artinux/creativity-ai-art-and-the-impact-on-the-diversity-of-cultural-expressions-7b62550d1cee">https://medium.com/artinux/creativity-ai-art-and-the-impact-on-the-diversity-of-cultural-expressions-7b62550d1cee</a></p><p><a href="https://www.deepneuron.org/lae-blogs/2023/3/31/chatgpt-in-education-jta5h#_ftn1">[6]</a> Artists as Workers: The Undoing of Yugoslav Socialism and the Politics of Unpaid Artistic Labor, Katja Praznik.</p><p><a href="https://www.deepneuron.org/lae-blogs/2023/3/31/chatgpt-in-education-jta5h#_ftnref1">[7]</a> <a href="https://yaledailynews.com/blog/2023/01/23/what-ai-art-means-for-society-according-to-yale-experts/">https://yaledailynews.com/blog/2023/01/23/what-ai-art-means-for-society-according-to-yale-experts/</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=a4b6e745a7b3" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>