<?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 Pat Inc on Medium]]></title>
        <description><![CDATA[Stories by Pat Inc on Medium]]></description>
        <link>https://medium.com/@PatisNLU?source=rss-6a381e209bf------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/2*gsyCrN1-Nl69eqfEXhPavw.png</url>
            <title>Stories by Pat Inc on Medium</title>
            <link>https://medium.com/@PatisNLU?source=rss-6a381e209bf------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Mon, 16 Nov 2026 06:23:10 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@PatisNLU/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[The Problem with AI State-Of-The-Art Methodology]]></title>
            <link>https://medium.datadriveninvestor.com/the-problem-with-ai-state-of-the-art-methodology-db30762d4b84?source=rss-6a381e209bf------2</link>
            <guid isPermaLink="false">https://medium.com/p/db30762d4b84</guid>
            <category><![CDATA[technology-trends]]></category>
            <category><![CDATA[chatbots]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[natural-language]]></category>
            <category><![CDATA[voice-assistant]]></category>
            <dc:creator><![CDATA[Pat Inc]]></dc:creator>
            <pubDate>Tue, 19 Feb 2019 03:21:59 GMT</pubDate>
            <atom:updated>2019-02-21T03:27:37.108Z</atom:updated>
            <content:encoded><![CDATA[<p><em>With our news feeds constantly filled with ‘machine beats human’ content, it can be hard to believe that the fundamental, hard AI problems haven’t been solved yet.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*vBTv0QsQZ6NdssiZ82ribQ.jpeg" /></figure><p>Last year, it was big news when AI created by Alibaba and Microsoft topped the SQuAD (Stanford Question and Answer Dataset) global reading comprehension test. The leaderboard on SQuAD’s website showed Alibaba’s and Microsoft’s scores as 82.44 and 82.65 respectively. The scores are higher than a human’s, which is 82.304.</p><p>But what does this mean for AI advancement? As <a href="https://www.theverge.com/2018/1/17/16900292/ai-reading-comprehension-machines-humans">James Vincent points out</a>, this was done ‘in a very specific and constrained task.’ Like a spreadsheet outperforms humans in calculations. However, it’s not the precursor to a singularity.</p><h4>AI today is good at narrow tasks, but not general. And this matters.</h4><p>Can we get good output and application from narrow tasks? Yes, we can. But does it mean the methodology and tools used for these narrow tasks, will be enough to take us to language understanding and remove the frustration we encounter daily with dumb chatbots, voice assistants and smart speakers? No.</p><blockquote>“Building a taller ladder does not get us to the moon.”</blockquote><p>If we pick the wrong target for benchmarking tests, we will miss the ‘right’ target, on every single go. One of the highly regarded Deep Learning fathers, <a href="https://www.technologyreview.com/s/612434/one-of-the-fathers-of-ai-is-worried-about-its-future/">Yoshua Bengio</a> said</p><blockquote>“I think we need to consider the hard challenges of AI and not be satisfied with short-term, incremental advances…we need to be able to extend it to do things like reasoning, learning causality, and exploring the world in order to learn and acquire information.”</blockquote><p>Autonomous vehicles have been driving millions of miles now, but are still only <a href="https://www.autonews.com/article/20180507/MOBILITY/180509837/texas-gets-self-driving-car-service-through-drive-ai">commercially used where the environment is tightly constrained</a>. There’s a problem here, and it’s not an engineering one. It’s a scientific one.</p><h4>Achieving full human language understanding and visual recognition is not easy.</h4><p>In order to understand a conversation, humans don’t need to have heard every conceivable word and phrase combination (which is infinite). We only need to know around 3,000 words, what’s needed for basic daily conversation.</p><p>Why do machines need billions of text corpora to try and understand? Or, drive millions of miles? Because machines can’t generalize. If the training data isn’t similar enough to the test data, the machine will fail.</p><p>As Gary Marcus says in <a href="https://syncedreview.com/2019/02/15/ai-hasnt-found-its-isaac-newton-gary-marcus-on-deep-learning-defects-frenemy-yann-lecun/">Synced</a>, ‘(deep learning) is very good at image classification, it’s very good at labelling what part of speech a word is. It’s not very good at understanding a scene or the meaning of a sentence. It’s not really good enough to drive a domestic robot around your home. There are lots of things it doesn’t do very well. There’s no AI that’s remotely as clever as my four-year-old or my six-year-old.”</p><p>In 2018 there was a steady stream of announcements from <a href="https://ai.googleblog.com/2018/05/duplex-ai-system-for-natural-conversation.html">Google Duplex</a>, <a href="https://allennlp.org/elmo">ELMo from Al2</a>, <a href="https://ai.googleblog.com/2018/11/open-sourcing-bert-state-of-art-pre.html">Google’s Bert</a>, and more recently, OpenAI’s <a href="https://blog.openai.com/better-language-models/">GPT2</a>.</p><p>The engineering back-up documentation is a watered-down version of the media portrayal. And, ensuing products like Bert’s predictive typeahead is curtailed after release, with pronoun restrictions.</p><p>Why? Because incremental improvements don’t lead to general and natural intelligence. Useful products can come out of technology, but a quick glance at Gartner’s hype cycle for 2018 shows the problem with inflated expectations for these tools and products. Deep Learning is at the peak of inflated expectations, and Virtual Assistants have turned the corner and are now on their journey toward the Trough of Disillusionment.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*Kq8g5MCR5tbPqbgX" /></figure><p>If 2018 was the year for ‘AI announcement hysteria’, it was also a year for questioning whether state-of-the-art methodologies can provide the way forward.</p><p>One of the field’s sharpest critics, Judea Pearl (Developer of Bayesian Networks) expressed concern in <a href="https://www.quantamagazine.org/to-build-truly-intelligent-machines-teach-them-cause-and-effect-20180515/">Quanta Magazine</a> that</p><blockquote>“the state-of-the-art in artificial intelligence today is merely a souped-up version of what machines could already do a generation ago: find hidden regularities in a large set of data…(and is) mired in probabilistic associations.”</blockquote><p>AI luminary, Geoff Hinton recently said in an <a href="https://www.axios.com/artificial-intelligence-pioneer-says-we-need-to-start-over-1513305524-f619efbd-9db0-4947-a9b2-7a4c310a28fe.html">Axios</a> interview “my view is to throw it all away and start again…I don’t think it’s how the brain works. We clearly don’t need all the labelled data.”</p><h4>So how do we progress?</h4><p>Pat’s solution to the fundamental, hard ‘AI’ problem of NLU is to engineer from science. And, that science is <a href="https://medium.com/pat-inc/patom-theory-explains-language-acquisition-part-1-609b3c42ed18">Patom Theory</a> and <a href="http://www.acsu.buffalo.edu/~rrgpage/rrg/RRGpaper.pdf">Role &amp; Reference Grammar</a> (RRG).</p><p>Cognitive Scientist and Pat’s Founder, <a href="https://en.wikipedia.org/wiki/John_Ball_(cognitive_scientist)">John Ball</a> spent years researching the problem of NLU, which lead him to Patom Theory. Studying how the brain approaches language, John hypothesised how to represent human language for a machine. In 2011, he came across the missing piece which was RRG, a similarly deep science focusing on how the world’s languages work, developed by <a href="http://arts-sciences.buffalo.edu/linguistics/faculty/robert-vanvalin.html">Robert D. Van Valin, Jr</a>. <strong>Like peanut butter meets chocolate.</strong></p><p>In 2016 Pat was founded, to exploit a scientific breakthrough for conversations between humans and machines.</p><p><a href="http://pat.ai/">Pat</a> deviates from traditional statistical NLP models and focuses on a brain-based approach to language, to correctly assign meaning to words in relation to the other words in context.</p><p>As a result, Pat is able to correctly analyze complex natural language interactions and directly connect phrases to meaning.</p><blockquote>Chatbot and voice assistants should be capable of full conversation, and provide useful, valuable and convenient customer experiences. In order to get there, we need to integrate science into engineering.</blockquote><h4><strong>To join Pat’s private beta, sign-up here: https://pat.ai/</strong></h4><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=db30762d4b84" width="1" height="1" alt=""><hr><p><a href="https://medium.datadriveninvestor.com/the-problem-with-ai-state-of-the-art-methodology-db30762d4b84">The Problem with AI State-Of-The-Art Methodology</a> was originally published in <a href="https://medium.datadriveninvestor.com/">DataDrivenInvestor</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>