<?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 Vilson Vieira on Medium]]></title>
        <description><![CDATA[Stories by Vilson Vieira on Medium]]></description>
        <link>https://medium.com/@aut0mata?source=rss-7a1781e5c866------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*KLOPbktRvPCsA7H2_Cl50Q.png</url>
            <title>Stories by Vilson Vieira on Medium</title>
            <link>https://medium.com/@aut0mata?source=rss-7a1781e5c866------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sat, 21 Nov 2026 15:40:10 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@aut0mata/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[Week in Review #0000]]></title>
            <link>https://aut0mata.medium.com/week-in-review-0000-e71bbb048049?source=rss-7a1781e5c866------2</link>
            <guid isPermaLink="false">https://medium.com/p/e71bbb048049</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[chatgpt]]></category>
            <category><![CDATA[langchain]]></category>
            <dc:creator><![CDATA[Vilson Vieira]]></dc:creator>
            <pubDate>Tue, 25 Apr 2023 19:28:58 GMT</pubDate>
            <atom:updated>2023-04-25T19:28:58.899Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*MqMIijzoHexEeFhKPtz4Zg.jpeg" /><figcaption>Prompt: “autonomous agents”</figcaption></figure><h3>2023–04–17 =&gt; 2023–04-21</h3><p>I’m trying to create a post per week to summaryze what I learned during the week, papers I read, news I found interesting, code I experimented and/or contributed to:</p><ul><li>Google Brain and DeepMind merged: it seems an answer to OpenAI and other independent AI labs, to consolidate an unique front for AI research on Google. Pretty curious about what’s coming next, given that Google and DeepMind never stopped releasing really important pieces that paved the way to current AI hype (e.g. Transformer, LLMs, etc).</li><li>Started exploring <a href="https://github.com/hwchase17/langchain/">LangChain</a>: I really like how they are organizing LLMs into Agents and external APIs as Tools, so we can chain them together to create LLM-based pipelines. They are really fast in incorporating new concepts (like Autonomous Agents), making it easier to explore most recent LLM pieces through only one API.</li><li>Started exploring Autonomous Agents: AutoGPT, BabyAGI, JARVIS, Generative Agents are getting great momentum: “Recursive” agents that start with one goal and keep exploring and increasing sub-tasks to reach the goal. When used together with tools like LangChain, can make pretty powerful tools.</li><li>Stability jumped on the LLM train by releasing its first LLM model.</li><li>It seems LLM models size and data must grow together (<a href="https://arxiv.org/abs/2203.15556">https://arxiv.org/abs/2203.15556</a>). Which means that scaling data is possibly the next step for LLM.</li><li>HuggingGPT (or <a href="https://github.com/microsoft/JARVIS">JARVIS</a>) was released and it’s pretty impressive: an LLM that can search and use ML models hosted on Hugging Spaces to solve tasks. It would be nice to have it integrated in LangChain.</li><li>Elon wants to sue MS and another of his rockets exploded, the launchpad seems OK, it only has a giant crater bellow it ;-)</li><li>Learnt about <a href="https://memprompt.com/">https://memprompt.com/</a></li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=e71bbb048049" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Quick survival guide to tmux]]></title>
            <link>https://aut0mata.medium.com/quick-survival-guide-to-tmux-166b8a4de6a8?source=rss-7a1781e5c866------2</link>
            <guid isPermaLink="false">https://medium.com/p/166b8a4de6a8</guid>
            <category><![CDATA[vim]]></category>
            <category><![CDATA[terminal]]></category>
            <category><![CDATA[linux]]></category>
            <category><![CDATA[tmux]]></category>
            <category><![CDATA[unix]]></category>
            <dc:creator><![CDATA[Vilson Vieira]]></dc:creator>
            <pubDate>Sat, 22 May 2021 11:51:20 GMT</pubDate>
            <atom:updated>2025-12-17T14:03:47.815Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*aovdly8gFj4PzoOJ7ehqRQ.png" /></figure><p>Yep, this is the one. You can stop searching now. Just bookmark this page and come back here when you wanna remember some basic <em>tmux</em> keystroke or to setup a basic <em>tmux</em> config in a fresh Linux/Unix box 😉</p><h4>🤔 Wait a minute, what the hell is tmux?</h4><p><em>Tmux</em> is a terminal multiplexer. Wait, what? Why I need that?</p><ul><li>You just connected to a Linux box through SSH, but hey, <strong>I want to run more than one command at a time</strong>! One shell is just not enough! Use <em>tmux</em> and you can have any number of terminals you want!</li><li>You connected to a Linux box through SSH, but hey, <strong>I want to run this command forever</strong>! Use <em>tmux</em> and you can run any command you want, detach from the terminal and it will keep running. <em>For-eh-vah</em>!</li><li>You love the good and old terminal, but <strong>you want more power to switch between shells really freaking fast</strong>. Use <em>tmux</em>!</li><li>You love both the terminal and <em>Vim</em>. Use <em>tmux</em>!</li></ul><h4>🤘 Cool! How I install it?</h4><p>At Ubuntu:</p><pre>sudo apt install tmux</pre><p>At Mac:</p><pre>brew install tmux</pre><p>Download this minimal config file I created at your home sweet home. It has the minimal required to improve your UX on tmux:</p><pre>wget <a href="https://gist.githubusercontent.com/automata/2a3e8a276b2fd05ae13d086422dc4140/raw/2bb7883e17427adb874ce14f60dc63f773460022/tmux.conf">https://gist.githubusercontent.com/automata/2a3e8a276b2fd05ae13d086422dc4140/raw/2bb7883e17427adb874ce14f60dc63f773460022/tmux.conf</a> -O ~/.tmux.conf</pre><h4>🤓 OK it was easy… how can I use it now?</h4><p>First, some basic concepts you&#39;ll become familiar with:</p><ul><li><em>Tmux</em> runs <strong>sessions</strong> where we can attach clients to</li><li>Each session can have many <strong>windows</strong> (listed on the bottom)</li><li>Each window can have many <strong>panes</strong> (the split panes you&#39;ll see)</li><li>Each new pane is a new <strong>bash</strong> prompt/terminal</li></ul><p>To <strong>create a new session</strong>:</p><pre>tmux</pre><p>Now you can type C-b dand you&#39;ll <strong>dettach</strong> from the session. It&#39;s still running there, but you can exit you&#39;re SSH session now and attach again later… and all your beloved terminals will still be running 😍</p><p>To <strong>attach to a session</strong> you just created:</p><pre>tmux a</pre><p>To <strong>list existing sessions</strong>:</p><pre>tmux ls</pre><p>To <strong>attach to a specific session</strong> (see the first column from previous command, that&#39;s the id you need to use here. I&#39;m using session zero):</p><pre>tmux a -t 0</pre><h4>🤸🏻‍♀ That&#39;s super cool! But how can I move around?!</h4><p>All commands start with C-b (C is the command key on Mac and thectrl key on Linux). So, basically always type the C-b <strong>prefix</strong> before almost every keystroke.</p><p>And here are the commands I use all the time and they should be enough for you to use <em>tmux</em> for weeks without having to worry much about it:</p><pre>C-b c                     Create a new window<br>C-b |                     Split window vertically<br>C-b -                     Split window horizontally<br>C-b &lt;arrow keys&gt;          Move between panes<br>C-b &lt;h,j,k,l&gt;             Move between panes Vim-style<br>C-b &lt;number&gt;              Go to window number n (listed at bottom)<br>C-b [                     Enter in &quot;scroll mode&quot;.<br>                          - Use arrows and page-down/up to move<br>                          - Type Enter to exit the mode.<br>C-b d                     Detach from session (to get back: tmux a)</pre><h4>😎 That&#39;s it?</h4><p>Yep… but nope 😉 <em>Tmux</em> is an awesome piece of software with many, many more possibilities. Check its official <a href="https://github.com/tmux/tmux">Git repos</a> for more and <a href="https://twitter.com/aut0mata">follow me on Twitter</a>, I&#39;ll keep sharing tools and tricks I use daily as a Machine Learning Engineer and Researcher that can probably help you too! 👋</p><p>🤩 <strong>UPDATE: Wait, there&#39;s more!</strong></p><p>From time to time I find more interesting commands, so I&#39;ll keep posting them here.</p><p><strong>Sending the same command to all panes at once:</strong></p><p>Let&#39;s say you have many panes open and you want to send the same commands to all of them:</p><pre>C-b : setw synchronize-panes on</pre><p>And to toggle it off:</p><pre>C-b : setw synchronize-panes off</pre><p><strong>Removing other attached client:</strong></p><p>You started your tmux and there&#39;s something wrong with the window size, it looks smaller than yours… maybe there&#39;s another client connected to your tmux running in a smaller window than yours! Just list the connected clients and remove the one you don&#39;t like:</p><pre>C-b D</pre><p><strong>My panes are not aligned:</strong></p><p>You can use the following command to iterate through many different layouts, maybe you&#39;re lucky and get your panes aligned as you wish:</p><pre>C-b &lt;space&gt;</pre><p><strong>My windows are numbered wrong:</strong></p><p>After a while, when you have removed windows in the middle, the numbers of your windows start to jump from n to n+x with x&gt;1 . To re-enumerate the windows, just use:</p><pre>C-b : move-window -r</pre><p><strong>I want to maximize one of the panes for a moment:</strong></p><p>When you have too many panes open, it starts to get pretty hard to read on those smaller windows. To maximize the pane you have active, just toggle maximize/minize:</p><pre>C-b z</pre><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=166b8a4de6a8" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[AI's eyes]]></title>
            <link>https://aut0mata.medium.com/ais-eyes-765b5d2c7c83?source=rss-7a1781e5c866------2</link>
            <guid isPermaLink="false">https://medium.com/p/765b5d2c7c83</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[computer-vision]]></category>
            <category><![CDATA[design]]></category>
            <category><![CDATA[machine-learning]]></category>
            <dc:creator><![CDATA[Vilson Vieira]]></dc:creator>
            <pubDate>Fri, 09 Apr 2021 02:52:46 GMT</pubDate>
            <atom:updated>2021-05-22T12:18:04.171Z</atom:updated>
            <content:encoded><![CDATA[<h4>How <a href="https://thegrid.io/">The Grid</a> was teaching its Artificial Intelligence to see and paint.</h4><p><em>Note: I wrote this post back in the day I worked as an AI Engineer at The Grid. It was a crazy ride, with lots of interesting people and challenges. I left The Grid years ago but some ideas are still relevant to augment design using AI. The content of this post were originally published on The Grid itself and since there&#39;s no mirrors around, I&#39;m sharing it here and I hope you enjoy it.</em></p><p>Augmenting and automating design of websites involve many steps and the areas I had my hands allover were on Computer Vision (CV) and Generative Art. We used a mix of traditional CV, generative algorithms and Machine Learning (ML) models, that when applied to design, got us the following features:</p><ul><li><strong>Saliency extraction</strong>: good to crop to region and text placement/overlay</li><li><strong>Face detection</strong>: for emotion recognition and supporting detection of salient regions</li><li><strong>Delaunay triangulation</strong>: to generate artistic backgrounds and picture styling</li><li><strong>Image classification</strong>: to extract information/semantics from pictures</li><li><strong>Sentiment analysis</strong>: to extract the general mood of a picture</li><li><strong>Color extraction</strong>: to extract/quantize colors from pictures</li></ul><p>Let&#39;s take a look on some of them in more details, shall we?</p><h3>Saliency extraction</h3><p>When we look into a picture we can naturally (and really fast!) determine what points get our attention first. A human face, a specific object, a region with higher contrast, those are both examples of <em>salient areas.</em> We can easily recognise those areas, but computers don&#39;t. So we have to teach them to.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*FsTzqKrj7tt00nkuzZ3d0g.jpeg" /><figcaption>Original image</figcaption></figure><p><a href="https://thegrid.io/">The Grid</a> used a <a href="https://github.com/automata/gmr-saliency">saliency detection algorithm</a> which analyses a given image looking for regions of high contrast. In summary, it divides the original image into small segments. Those segments are analysed and just the most contrasting ones are selected. We call them <em>salient regions </em>(as shown in the side figure).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*frdfPWtvOiKjXvyxXBTjXQ.png" /><figcaption>Salient regions extracted from the image. Higher gray levels indicate more salient areas.</figcaption></figure><p>Knowing which regions are most salient we can now crop the original image to show only those regions — to fit on any screen size, for example. Or we can place text on top of images avoiding the salient regions.</p><h3>Generative images and image processing</h3><p>Besides analytics, we can use extracted information from text, images or videos to synthesise new content. We can generate an infinite number of unique images using combinations of different shapes, curves and colours.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/208/1*AANFvkfmJJ2POLmFi62jXg.gif" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/208/1*cd2R3nJAXDyy6qWyKtfT3A.gif" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/208/1*PtacjDuGVVxgw28px1X90w.gif" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/208/1*7NZamW06czDa_m0-ogvl4g.gif" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/208/1*xYHu8ltqLjPGVCS28vVAGQ.gif" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/208/1*jpOLoquKHxvEvP_cyZE0eQ.gif" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/208/1*CbsuppwF8FKb__GOfgsOmA.gif" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/208/1*iUZSlGMgG2hik8FjD0_utg.gif" /></figure><p>We can also apply more complex algorithms like Delaunay Triangulation to obtain more interesting and unique meshes for <em>favbanners</em> or backgrounds.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/730/1*8L-jnDPFj6H2Yllam57IPQ.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/730/1*TvjdYFD3ZnlWQvzkmVI_XQ.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/730/1*u_peAgpYjHiyrPhVx-HFqA.png" /><figcaption>Generative images by Delaunay Triangulation</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/730/1*rP7EHOQiwUHfY19LCZ8tFQ.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/730/1*rb2nnJc4JqS0NZtY-6ryRw.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/730/1*ONc58Or_PnXrjBkAuzcmDA.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/752/1*8r7K1YwZ5R3Lpf__dPT7-Q.png" /></figure><p>Or we can go beyond and use the same algorithm to style existing pictures.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*aCd6lHpMflMr1ZSNU6rLDg.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*BeTzY4BP65-90XKDZ0-e7Q.png" /><figcaption>Image styling using Delaunay Triangulation</figcaption></figure><p>Using our open-source image processing pipeline <a href="https://github.com/jonnor/imgflo">imgflo</a> we were able to manipulate images. It was possible to create any image filter and apply them to enhance pictures.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/150/1*voifTMbtZFSeOf4IDZ3HBA.jpeg" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/150/1*rfkvDFWNx9CoVSlNuwr4mg.jpeg" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/150/1*Ojb6cSTc2QcmXhL2vJuSkQ.jpeg" /><figcaption>Filters created on <a href="https://github.com/jonnor/imgflo">imgflo</a> being applied to the same image. Original picture by <a href="https://flic.kr/p/71KH5k">Sharon Mollerus</a>.</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/150/1*x48k9t1oPxBQ6sn6Qrj2ag.jpeg" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/150/1*j8-Gjff7Jl6bWFjSz4ArmA.jpeg" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/150/1*Sth27O4QuqwPFKtEMhRynQ.jpeg" /></figure><p>Imgflo is a runtime for Flowhub, a visual programming environment that made it simple to build image processing graphs.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/822/1*VjOvPnyheQFKQCoxJ33IwA.png" /><figcaption>A Flowhub graph to image processing on imgflo.</figcaption></figure><h4>What&#39;s next?</h4><p>It&#39;s more than 6 years since I wrote this post. Deep Learning happened, Computer Vision was completely changed by it and other ML methods. However, the fundamental problems on design are still there and we can leverage a lot from ML/CV to augment and automate graphical and web design.</p><p>If you want to know more about my last developments on blending AI/ML with design and creativity, <a href="https://twitter.com/aut0mata">follow me on Twitter</a>!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=765b5d2c7c83" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Notes and Comments on Yann LeCun Interview with Lex Fridman]]></title>
            <link>https://aut0mata.medium.com/notes-and-comments-on-yann-lecun-interview-with-lex-fridman-f8cfe7d88db0?source=rss-7a1781e5c866------2</link>
            <guid isPermaLink="false">https://medium.com/p/f8cfe7d88db0</guid>
            <category><![CDATA[lex-friedman]]></category>
            <category><![CDATA[coding]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[lecun]]></category>
            <category><![CDATA[machine-learning]]></category>
            <dc:creator><![CDATA[Vilson Vieira]]></dc:creator>
            <pubDate>Fri, 09 Apr 2021 02:19:53 GMT</pubDate>
            <atom:updated>2021-04-09T02:19:53.001Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*EuxSNnPBBhKR9JW7rXvciw.jpeg" /></figure><p>While watching a <a href="https://www.youtube.com/watch?v=SGSOCuByo24">interview</a> with Yann LeCun I took some notes and written down some comments of my own, together with useful references to works cited by LeCun. Other people <a href="https://www.facebook.com/pyxelr/posts/2501327949927255">shared valuable</a> <a href="https://twitter.com/DCasBol/status/1169663286327107585">notes already</a> and here it goes my free interpretations. They can run wild sometimes and are probably wrong or open to debate, so please feel free to let me know if I misunderstood some concepts.</p><blockquote>Do not believe on well defined logic/symbolic rules, but on gradient-based learning.</blockquote><blockquote>Logic representation and graphs are too rigid and too brittle/fragile.</blockquote><blockquote>Main issue is knowledge acquisition: how to reduce a bunch of data into a graph of this type? Relies on experts to create those rules.</blockquote><blockquote>Representing knowledge with symbols and manipulating symbols with logic is incompatible with learning</blockquote><p>We faced those problems already at <a href="https://extend.ai/">Extend AI</a> (and previously at The Grid) while applying Constraint Programming to tasks like 2D layout of web sites or floor plan design. It&#39;s really hard, time consuming and unpractical to depend on humans to define formal rules or constraints. However, it seems to exist room to train ML models to define constraints (like in autonomous Probabilistic Programming Languages) from data.</p><blockquote>Geoffrey Hinton advocates replacing symbols by vectors of probabilities and logic by continuous functions as a way to make it compatible with learning</blockquote><blockquote>León Bottou: a reasoning system should be able to manipulate objects in a space and result in objects in the same space, so it can store it again on the same memory space: <a href="https://arxiv.org/pdf/1102.1808.pdf">https://arxiv.org/pdf/1102.1808.pdf</a></blockquote><p>Brings the notion of self-programmable models and if we force ourselves to build up on Neural Networks structures, Connectionism can probably embrace/simulate Symbolism. Maybe here the notion of <a href="https://medium.com/@karpathy/software-2-0-a64152b37c35">Software 2.0</a> can really shine and emerge from.</p><blockquote>ML is the science of sloppiness :-)</blockquote><blockquote>Neural Networks are capable of reasoning, intuition is based on the fact that we already have one that works: our brain ;-)</blockquote><blockquote>We need a memory “device” to store large amount of factual memories during a certain period of time.</blockquote><blockquote>Three types of memory:</blockquote><blockquote>1. On visual cortex, only stores state for about 20 seconds and then vanishes</blockquote><blockquote>2. On hippocampus, short term memory, eg. partial map of current place, last sentences somebody said minutes ago.</blockquote><blockquote>3. On synapses, long term memory</blockquote><blockquote>We’re interested in hippocampus as memory-like device.</blockquote><p>Deep Nets with some kind of memory mechanism like — LSTM, Attention or Transformer — seem trying to recreate the hippocampus at some level. In this way a model can get data from memory, crunch on it and store it back, repeat. Chains of facts can be recursive and we need to iterate through it, updating knowledge/memory.</p><blockquote>Let’s get a simulation, a game… if we can simulate almost everything in this closed system we solved the data generation problem, not the “blurry” predictions.</blockquote><blockquote>We can generate data, OK, however, some unpredictable situation/event can happen, and the resulting prediction of model will be blurry prediction (i.e. all possible states/weights it could predict, so blurry). Something is missing, how to deal with those blurry predictions?</blockquote><blockquote>Hypothesis: model-based reinforcement learning. We can have some predictive models trained on more general (or specialized?) contexts like simple physics so we can “ask” for those models help to un-blurry the prediction.</blockquote><p>Sounds similar to<a href="https://youtu.be/7ROelYvo8f0"> what prof. Josh Tenenbaum says</a> about having simple physics simulation engine mimicking reality, so we can understand in advance how some events should be, or what happens if we perform some actions.</p><p>However it&#39;s hard to know which and how many are those “primitive” models.</p><blockquote>RL should not just be stupid like training 200 humans years a player to beat Starcraft. Imagine how we learn to drive a car. We already learned (as a predictive model, when we were 8–9 months old) that things can fall of tables, so when we were learning to drive we didn’t push ourselves out of cliffs because we already know that the car would fall out the cliff the same way block pieces fallen from tables when we were babies.</blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*2qb7C1Fo5q-IRy-MIJ_IQQ.png" /><figcaption>Emmanuel Dupoux: observations of baby concepts acquisition during months of life. Source: <a href="https://www.slideshare.net/rouyunpan/deep-learning-hardware-past-present-future">https://www.slideshare.net/rouyunpan/deep-learning-hardware-past-present-future</a></figcaption></figure><blockquote>We need RL systems with such predictive models that allow us to avoid doing stupid things.</blockquote><blockquote>For driving a car, we still need imitation learning because we’re learning from some instructor, but most of the time it’s all about learning a world model (mechanical physics, for instance). And such predictive models can be transferred from scene to scene because stupid things happens everywhere.</blockquote><blockquote>So the main question is: how do we learn models of the world?! That’s what self-supervised learning is trying to help.</blockquote><p>A basic Self-supervised Learning model in that context could take some amount of labeled and unlabeled data, train a supervised model on labeled data and run it on both labeled/unlabeled samples. If average loss drops, the new samples can be part of labeled data now.</p><p>That brings the question of how to deal with false positives because they could be taken as valid training data, maybe that&#39;s why self-supervised models are commonly applied to sequential data in “fill the blank” problems.</p><blockquote>An Autonomous Intelligent System (with possibly human level intelligence) should have three pieces:</blockquote><blockquote>1. World Model: update-able, self-supervised learnable.</blockquote><blockquote>2. Objective Function: based on <a href="https://en.wikipedia.org/wiki/Basal_ganglia">basal ganglia</a> (level of contentment).</blockquote><blockquote>3. Predictive System / Policy Maker: observe world at time t; what happens to the world if I take this action; collect many answers represented with some uncertainty level (distributions?); run model for some situations/hypothesis; min/max objective function (by energy minimization?). The goal is to answer: what sequence of actions I need to take to increase contentment level.</blockquote><p>To be really autonomous, do such world models need to be able to program themselves? Interesting that emotions like fear and happiness become protagonists on guiding actions.</p><blockquote>Question to an human-level intelligent system: what’s the wind? Needs to understand world observation, adapt its model to corrected one and show some level of causality reasoning</blockquote><blockquote>No need for embodiment but grounding.</blockquote><p>It seems some road to human-level intelligence would be a <strong>model-based Reinforcement Learning</strong> system running on a <strong>world simulation</strong> with models being trained as <strong>Self-supervised Learning</strong> models.</p><p>We&#39;re only scratching the surface of such models, but the journey ahead is fascinating! Please <a href="https://twitter.com/aut0mata">follow me on Twitter</a> where I share more about ML and AI in general!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f8cfe7d88db0" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>