2023-01-05-the-mechanistic-fallacy-and-modelling-how-we-think-caught-by-surprise.html

---
title: "Caught by Surprise Retrospective: The Mechanistic Fallacy and
Modelling How We Think"
venue: "ELLIS Unconference"
abstract: "<p>In this talk, I revisit a talk given at NeurIPS 2015 where
I speculated about the next directions for ML. The aim is to find
surprising things, so I’ll try to reconstruct what I was thinking at the
time and compare that to what I think now.</p>
<p>In this talk we will discuss how our current set of modelling
solutions relates to dual process models from psychology. By analogising
with layered models of networks we first address the danger of focussing
purely on mechanism (or biological plausibility) when discussion
modelling in the brain. We term this idea the mechanistic fallacy. In an
attempt to operate at a higher level of abstraction, we then take a
conceptual approach and attempt to map the broader domain of mechanistic
and phenomological models to dual process ideas from psychology. it
seems that System 1 is closer to phenomological and System 2 is closer
to mechanistic ideas. We will draw connections to surrogate modelling
(also known as emmulation) and speculate that one role of System 2 may
be to provide additional simulation data for System 1.</p>"
author:
- given: Neil D.
  family: Lawrence
  url: http://inverseprobability.com
  institute: University of Sheffield
  twitter: lawrennd
  gscholar: r3SJcvoAAAAJ
  orcid: 
edit_url: https://github.com/lawrennd/talks/edit/gh-pages/_cognitive-science/the-mechanistic-fallacy-and-modelling-how-we-think-caught-by-surprise.md
date: 2023-01-05
published: 2023-01-05
week: 0
reveal: 2023-01-05-the-mechanistic-fallacy-and-modelling-how-we-think-caught-by-surprise.slides.html
transition: None
layout: talk
categories:
- notes
---



<!-- Do not edit this file locally. -->
<!---->
<!-- Do not edit this file locally. -->
<!-- Do not edit this file locally. -->
<!-- The last names to be defined. Should be defined entirely in terms of macros from above-->
<!--

-->
<h1 id="context">Context</h1>
<p>The background to this talk is a talk I gave in 2015 at NeurIPS in
Montreal called “Statistical Methods for Understanding Neural Systems”.
The workshop was organised by Allie Fletcher, Jakob Macke, Ryan P.
Adams, Jascha Sohl-Dickstein, there’s a summar here: <a
href="https://memming.wordpress.com/2015/12/15/nips-2015-workshops/"
class="uri">https://memming.wordpress.com/2015/12/15/nips-2015-workshops/</a>.</p>
<p>My sense at the time was that the “rise of the neural network” was an
opportunity to reinject some of the “funkiness” into the field. In
particular, to revisit ideas from neuroscience and cognitive science and
work on algorithms that are inspired by those ideas.</p>
<p>With that in mind I agreed to give a talk when Allie invited me.</p>
<p>This photo is from the panel session.</p>
<div class="figure">
<div id="panel-statistical-neural-figure" class="figure-frame">
<div class="centered centered" style="">
<img class="" src="https://inverseprobability.com/slides/diagrams//people/2015-12-11-17-51-37-1.jpg" width="80%" height="auto" align="center" style="background:none; border:none; box-shadow:none; display:block; margin-left:auto; margin-right:auto;vertical-align:middle">
</div>
</div>
<div id="panel-statistical-neural-magnify" class="magnify"
onclick="magnifyFigure(&#39;panel-statistical-neural&#39;)">
<img class="img-button" src="{{ '/assets/images/Magnify_Large.svg' | relative_url }}" style="width:1.5ex">
</div>
<div id="panel-statistical-neural-caption" class="caption-frame">
<p>Figure: Panel session at the Statistical Methods for Understanding
Neural Systems workshop.</p>
</div>
</div>
<p>It was a great workshop and with retrospect, I now think of it as the
last event at NeurIPS that I attended that captured the spirit of those
early meetings. It was in a small room, which was packed, and had a
diversity of people from different fields, offering opinions in an open
and constructive manner.</p>
<p>At the time the Human Brain project was recently launched, and I
think it’s fair to say that most people in the room felt it was a
colossal waste of money. The objective was to build a full simulation of
the brain, and the counter point was, well even if you could do that,
what do you learn by it.</p>
<div class="figure">
<div id="hbp-logo-figure" class="figure-frame">
<div class="centered centered" style="">
<img class="" src="https://inverseprobability.com/slides/diagrams//ai/human-brain-project.jpg" width="80%" height="auto" align="center" style="background:none; border:none; box-shadow:none; display:block; margin-left:auto; margin-right:auto;vertical-align:middle">
</div>
</div>
<div id="hbp-logo-magnify" class="magnify"
onclick="magnifyFigure(&#39;hbp-logo&#39;)">
<img class="img-button" src="{{ '/assets/images/Magnify_Large.svg' | relative_url }}" style="width:1.5ex">
</div>
<div id="hbp-logo-caption" class="caption-frame">
<p>Figure: Logo of the human brain project.</p>
</div>
</div>
<p>So by starting with the HBP logo, I was highlighting that even a
“faithful” mechanistic model isn’t what we actually want, because it
doesn’t enable us to understand the system. It’s just a mirror of the
existing system.</p>
<p>Bernhard and I often talk about the issue of the kingdom that made
maps that were more and more accurate. It features in a Jorge Luis
Borges story, but predating that it appears in a Lewis Carroll
story.</p>
<blockquote>
<p>“What a useful thing a pocket-map is!” I remarked.</p>
<p>“That’s another thing we’ve learned from your Nation,” said Mein
Herr, “map-making. But we’ve carried it much further than you. What do
you consider the largest map that would be really useful?”</p>
<p>“About six inches to the mile.”</p>
<p>“Only six inches!” exclaimed Mein Herr. “We very soon got to six
yards to the mile. Then we tried a hundred yards to the mile. And then
came the grandest idea of all ! We actually made a map of the country,
on the scale of a mile to the mile!”</p>
<p>“Have you used it much?” I enquired.</p>
<p>“It has never been spread out, yet,” said Mein Herr: “the farmers
objected: they said it would cover the whole country, and shut out the
sunlight ! So we now use the country itself, as its own map, and I
assure you it does nearly as well.”</p>
<p>from Lewis Carroll, Sylvie and Bruno Concluded, Chapter XI, London,
1895</p>
</blockquote>
<div class="figure">
<div id="rube-goldberg-figure" class="figure-frame">
<div class="centered" style="">
<img class="" src="https://inverseprobability.com/slides/diagrams//physics/rube-goldbergs-self-operating-napkin.gif" width="70%" height="auto" align="center" style="background:none; border:none; box-shadow:none; display:block; margin-left:auto; margin-right:auto;vertical-align:middle">
</div>
</div>
<div id="rube-goldberg-magnify" class="magnify"
onclick="magnifyFigure(&#39;rube-goldberg&#39;)">
<img class="img-button" src="{{ '/assets/images/Magnify_Large.svg' | relative_url }}" style="width:1.5ex">
</div>
<div id="rube-goldberg-caption" class="caption-frame">
<p>Figure: Rube Goldberg’s self operating napkin.</p>
</div>
</div>
<p>Also of interest to me was general properties of models, how we think
about complex objects and I used a layered model of the computer network
to represent the contextual idea of a model. That different layers are
different ways of talking about the system (equivalent to different
maps).</p>
<div class="figure">
<div id="ip-stack-connections-figure" class="figure-frame">
<object class="svgplot " data="https://inverseprobability.com/slides/diagrams//compsci/IP_stack_connections.svg" width="60%" style=" ">
</object>
</div>
<div id="ip-stack-connections-magnify" class="magnify"
onclick="magnifyFigure(&#39;ip-stack-connections&#39;)">
<img class="img-button" src="{{ '/assets/images/Magnify_Large.svg' | relative_url }}" style="width:1.5ex">
</div>
<div id="ip-stack-connections-caption" class="caption-frame">
<p>Figure: Network Layer models in Computer Science.</p>
</div>
</div>
<div class="figure">
<div id="mouse-47172-figure" class="figure-frame">
<object class="svgplot " data="https://inverseprobability.com/slides/diagrams//mouse-47172.svg" width="80%" style=" ">
</object>
</div>
<div id="mouse-47172-magnify" class="magnify"
onclick="magnifyFigure(&#39;mouse-47172&#39;)">
<img class="img-button" src="{{ '/assets/images/Magnify_Large.svg' | relative_url }}" style="width:1.5ex">
</div>
<div id="mouse-47172-caption" class="caption-frame">
<p>Figure: The mouse is a model, a physical model of how the world
works. This drawing is a drawing of the mouse, so it’s a model of the
model.</p>
</div>
</div>
<p>The point about these stories was to highlight something we talked
about yesterday, that the model is contextual, like a good map. It needs
to be something that we can use to answer a specific question. And like
the different map scales are useful for walkers, cyclist and car
drivers. And the different features on the map are also appropriate for
different users, the right form of the model varies according to the
user, or more broadly the context in which the user is operating.</p>
<p>I went on to argue that the rise of machine learning was being driven
by the rise in data availability, as depicted in the graph below which
shows digital data availability up to 2007. I suspect that to draw such
a graph today would render the analogue component of data too small to
fit.</p>
<div class="figure">
<div id="hilbert-info-growth-figure" class="figure-frame">
<object class="svgplot " data="https://inverseprobability.com/slides/diagrams//data-science/hilbert-info-growth.svg" width="80%" style=" ">
</object>
</div>
<div id="hilbert-info-growth-magnify" class="magnify"
onclick="magnifyFigure(&#39;hilbert-info-growth&#39;)">
<img class="img-button" src="{{ '/assets/images/Magnify_Large.svg' | relative_url }}" style="width:1.5ex">
</div>
<div id="hilbert-info-growth-caption" class="caption-frame">
<p>Figure: The Global storage capacity between 1986 and 2007 <span
class="citation" data-cites="Hilbert:information11">Hilbert and López
(2011)</span></p>
</div>
</div>
<div class="figure">
<div id="big-data-figure" class="figure-frame">
<div class="centered" style="">
<img class="" src="https://inverseprobability.com/slides/diagrams//ai/big-data.png" width="70%" height="auto" align="center" style="background:none; border:none; box-shadow:none; display:block; margin-left:auto; margin-right:auto;vertical-align:middle">
</div>
</div>
<div id="big-data-magnify" class="magnify"
onclick="magnifyFigure(&#39;big-data&#39;)">
<img class="img-button" src="{{ '/assets/images/Magnify_Large.svg' | relative_url }}" style="width:1.5ex">
</div>
<div id="big-data-caption" class="caption-frame">
<p>Figure: Word cloud for big data.</p>
</div>
</div>
<p>One of the interesting features about machine learning is how it
includes relatively simple (what I call weakly mechanistic) assumptions
about the data. I think the interesting aspect of machine learning
algorithms is the extent to w hich they allow the data to speak. For
example the convolutional neural networks encode translation
invariance.</p>
<div class="figure">
<div id="htm-hierarchy-figure" class="figure-frame">
<div class="centered" style="">
<img class="" src="https://inverseprobability.com/slides/diagrams//ai/htm-hierarchy-example.png" width="80%" height="auto" align="center" style="background:none; border:none; box-shadow:none; display:block; margin-left:auto; margin-right:auto;vertical-align:middle">
</div>
</div>
<div id="htm-hierarchy-magnify" class="magnify"
onclick="magnifyFigure(&#39;htm-hierarchy&#39;)">
<img class="img-button" src="{{ '/assets/images/Magnify_Large.svg' | relative_url }}" style="width:1.5ex">
</div>
<div id="htm-hierarchy-caption" class="caption-frame">
<p>Figure: Hierarchical models for machine learning.</p>
</div>
</div>
<p>An additional model for intelligence comes from logic, which is often
represented by Holmes and Watson, I could have been making one of two
points here, and I don’t recall which it was. Either I was using the
picture to remind me about graphical models, which are a very
interesting way of introducing conditional independencies, or I was
being rude about first order logic as a route to intelligence.</p>
<div class="figure">
<div id="holmes-watson-figure" class="figure-frame">
<div class="centered centered" style="">
<img class="" src="https://inverseprobability.com/slides/diagrams//ai/holmes-watson.jpg" width="80%" height="auto" align="center" style="background:none; border:none; box-shadow:none; display:block; margin-left:auto; margin-right:auto;vertical-align:middle">
</div>
</div>
<div id="holmes-watson-magnify" class="magnify"
onclick="magnifyFigure(&#39;holmes-watson&#39;)">
<img class="img-button" src="{{ '/assets/images/Magnify_Large.svg' | relative_url }}" style="width:1.5ex">
</div>
<div id="holmes-watson-caption" class="caption-frame">
<p>Figure: Holmes and Watson discussing a case.</p>
</div>
</div>
<p>One of the limitations of simplistic models of thought is nicely
revealed by the “trolley problem” tropes, where human ethics is judged
by asking for decisions around whether to take actions where a runaway
trolley can either be diverted through some points, or by “throwing a
fat man off a bridge”.</p>
<h2 id="the-trolley-problem">The Trolley Problem</h2>
<div style="text-align:right">
<span class="editsection-bracket" style="">[</span><span
class="editsection"
style=""><a href="https://github.com/lawrennd/snippets/edit/main/_philosophy/includes/trolley-switch.md" target="_blank" onclick="ga('send', 'event', 'Edit Page', 'Edit', 'https://github.com/lawrennd/snippets/edit/main/_philosophy/includes/trolley-switch.md', 13);">edit</a></span><span class="editsection-bracket" style="">]</span>
</div>
<div class="figure">
<div id="trolley-problem-original-figure" class="figure-frame">
<div class="centered centered" style="">
<img class="negate" src="https://inverseprobability.com/slides/diagrams//ai/Trolley_1.jpg" width="80%" height="auto" align="center" style="background:none; border:none; box-shadow:none; display:block; margin-left:auto; margin-right:auto;vertical-align:middle">
</div>
</div>
<div id="trolley-problem-original-magnify" class="magnify"
onclick="magnifyFigure(&#39;trolley-problem-original&#39;)">
<img class="img-button" src="{{ '/assets/images/Magnify_Large.svg' | relative_url }}" style="width:1.5ex">
</div>
<div id="trolley-problem-original-caption" class="caption-frame">
<p>Figure: The trolley problem in its original form.</p>
</div>
</div>
<h2 id="the-push-and-the-trolley">The Push and the Trolley</h2>
<div style="text-align:right">
<span class="editsection-bracket" style="">[</span><span
class="editsection"
style=""><a href="https://github.com/lawrennd/snippets/edit/main/_philosophy/includes/trolley-push.md" target="_blank" onclick="ga('send', 'event', 'Edit Page', 'Edit', 'https://github.com/lawrennd/snippets/edit/main/_philosophy/includes/trolley-push.md', 13);">edit</a></span><span class="editsection-bracket" style="">]</span>
</div>
<div class="figure">
<div id="the-trolley-problem-figure" class="figure-frame">
<div class="centered centered" style="">
<img class="negate" src="https://inverseprobability.com/slides/diagrams//ai/trolley2.jpg" width="80%" height="auto" align="center" style="background:none; border:none; box-shadow:none; display:block; margin-left:auto; margin-right:auto;vertical-align:middle">
</div>
</div>
<div id="the-trolley-problem-magnify" class="magnify"
onclick="magnifyFigure(&#39;the-trolley-problem&#39;)">
<img class="img-button" src="{{ '/assets/images/Magnify_Large.svg' | relative_url }}" style="width:1.5ex">
</div>
<div id="the-trolley-problem-caption" class="caption-frame">
<p>Figure: The trolley problem.</p>
</div>
</div>
<p>I later explored more ideas on these problems and the need for
diversity in decision making blog post on <a
href="http://inverseprobability.com/2017/11/15/decision-making">in this
blog post</a>.</p>
<p>My memory is that I shifted to use the rest of the talk to reveal my
secret interest in cognitive science, and was sharing some rough ideas
from popular books and how these facets might map back to the way we
perceive the world.</p>
<p>One of my major interests at the time (and in an ongoing way) was the
manner in which the machines we were building were not <em>sentient</em>
in the manner of the systems that we were being warned about by Elon
Musk, Stephen Hawking, Nick Bostrom and others, they were non sentient.
And therein lied the problem. I described these systems as “System
Zero”, taking my cue from dual-process models in psychology which
suggest that we have two thinking processes … fast and slow.</p>
<h3 id="the-elephant-and-its-rider">The Elephant and its Rider</h3>
<div style="text-align:right">
<span class="editsection-bracket" style="">[</span><span
class="editsection"
style=""><a href="https://github.com/lawrennd/snippets/edit/main/_ai/includes/elephant-boy-poster.md" target="_blank" onclick="ga('send', 'event', 'Edit Page', 'Edit', 'https://github.com/lawrennd/snippets/edit/main/_ai/includes/elephant-boy-poster.md', 13);">edit</a></span><span class="editsection-bracket" style="">]</span>
</div>
<div class="figure">
<div id="elephant-boy-poster-figure" class="figure-frame">
<div class="centered centered" style="">
<img class="" src="https://inverseprobability.com/slides/diagrams//ai/elephant-boy-poster.jpg" width="40%" height="auto" align="center" style="background:none; border:none; box-shadow:none; display:block; margin-left:auto; margin-right:auto;vertical-align:middle">
</div>
</div>
<div id="elephant-boy-poster-magnify" class="magnify"
onclick="magnifyFigure(&#39;elephant-boy-poster&#39;)">
<img class="img-button" src="{{ '/assets/images/Magnify_Large.svg' | relative_url }}" style="width:1.5ex">
</div>
<div id="elephant-boy-poster-caption" class="caption-frame">
<p>Figure: The elephant and its rider is an analogy used by Haidt to
describe the difference between System One and System Two.</p>
</div>
</div>
<p>At the time one of my favourite approaches for thinking about these
systems was from Jonathan Haidt’s book, where he suggests that the fast
thinking side is an elephant, and the slow is a rider.</p>
<h3 id="the-righteous-mind">The Righteous Mind</h3>
<div style="text-align:right">
<span class="editsection-bracket" style="">[</span><span
class="editsection"
style=""><a href="https://github.com/lawrennd/snippets/edit/main/_ai/includes/the-righteous-mind.md" target="_blank" onclick="ga('send', 'event', 'Edit Page', 'Edit', 'https://github.com/lawrennd/snippets/edit/main/_ai/includes/the-righteous-mind.md', 13);">edit</a></span><span class="editsection-bracket" style="">]</span>
</div>
<div class="figure">
<div id="the-righteous-mind-figure" class="figure-frame">
<div class="centered" style="">
<img class="" src="https://inverseprobability.com/slides/diagrams//ai/the-righteous-mind.png" width="40%" height="auto" align="center" style="background:none; border:none; box-shadow:none; display:block; margin-left:auto; margin-right:auto;vertical-align:middle">
</div>
</div>
<div id="the-righteous-mind-magnify" class="magnify"
onclick="magnifyFigure(&#39;the-righteous-mind&#39;)">
<img class="img-button" src="{{ '/assets/images/Magnify_Large.svg' | relative_url }}" style="width:1.5ex">
</div>
<div id="the-righteous-mind-caption" class="caption-frame">
<p>Figure: The Righteous Mind, by Jonathan Haidt, a social psychologist,
uses the analogy of the elephant and the rider for dual process
cognition.</p>
</div>
</div>
<p>Another book on these models is from the sports psychologist, Steve
Peters, who was a former colleague from the University of Sheffield.</p>
<h1 id="the-chimp-paradox">The Chimp Paradox</h1>
<div style="text-align:right">
<span class="editsection-bracket" style="">[</span><span
class="editsection"
style=""><a href="https://github.com/lawrennd/snippets/edit/main/_books/includes/the-chimp-paradox.md" target="_blank" onclick="ga('send', 'event', 'Edit Page', 'Edit', 'https://github.com/lawrennd/snippets/edit/main/_books/includes/the-chimp-paradox.md', 13);">edit</a></span><span class="editsection-bracket" style="">]</span>
</div>
<div class="figure">
<div id="chimp-paradox-figure" class="figure-frame">
<div class="centered centered" style="">
<img class="" src="https://inverseprobability.com/slides/diagrams//ai/the-chimp-paradox.jpg" width="30%" height="auto" align="center" style="background:none; border:none; box-shadow:none; display:block; margin-left:auto; margin-right:auto;vertical-align:middle">
</div>
</div>
<div id="chimp-paradox-magnify" class="magnify"
onclick="magnifyFigure(&#39;chimp-paradox&#39;)">
<img class="img-button" src="{{ '/assets/images/Magnify_Large.svg' | relative_url }}" style="width:1.5ex">
</div>
<div id="chimp-paradox-caption" class="caption-frame">
<p>Figure: The Chimp Paradox is a variation of the dual process model
used by the sports psychologist Steve Peters.</p>
</div>
</div>
<p>And I attended a talk of his where he suggested that these were just
analogies that were highlighting work from early psychoanalysis such a
Freud’s model of the Id, the Ego and the Superego, showing the long
history of the dual process models (equivalent to the Chimp, the Human
and the Computer in Peters’s model).</p>
<div class="figure">
<div id="id-ego-superego-figure" class="figure-frame">
<div class="centered centered" style="">
<img class="" src="https://inverseprobability.com/slides/diagrams//cognitive-science/id-ego-superego.jpg" width="80%" height="auto" align="center" style="background:none; border:none; box-shadow:none; display:block; margin-left:auto; margin-right:auto;vertical-align:middle">
</div>
</div>
<div id="id-ego-superego-magnify" class="magnify"
onclick="magnifyFigure(&#39;id-ego-superego&#39;)">
<img class="img-button" src="{{ '/assets/images/Magnify_Large.svg' | relative_url }}" style="width:1.5ex">
</div>
<div id="id-ego-superego-caption" class="caption-frame">
<p>Figure: Freud’s model of the Id, the Ego and the Superego.</p>
</div>
</div>
<p>My suggestion was that the AI that we’d built was operating as a
System Zero that was interfacing with our fast thinking, potentially to
the detriment of our autonomy. There’s a blog post on <a
href="http://inverseprobability.com/2015/12/04/what-kind-of-ai/">blog
post here where you can read more</a>.</p>
<p>That process was being driven by the extent to which we were (and
are) using data to create these “non-sentient” systems that were making
decisions about us. From an ethical perspective this seemed problematic
and our work on the data trusts initiative <a
href="https://datatrusts.uk" class="uri">https://datatrusts.uk</a> was
inspired by these concerns. Currently we are supporting three pilot
projects which attempt to give control of data back to the “citizens”.
See also this <a
href="https://www.theguardian.com/media-network/2016/jun/03/data-trusts-privacy-fears-feudalism-democracy">Guardian
article</a>.</p>
<p>Other work I’m excited about that responds to these problems is
Sylvie Delacroix’s book on <a
href="https://www.bloomsbury.com/uk/habitual-ethics-9781509920419/">habitual
ethics</a>. See also the youtube video below.</p>
<div class="figure">
<div id="habitual-ethics-figure" class="figure-frame">
<iframe width="640" height="480" src="https://www.youtube.com/embed/OHweOGW5Ju4?start=" frameborder="0" allow="autoplay; encrypted-media" allowfullscreen>
</iframe>
</div>
<div id="habitual-ethics-magnify" class="magnify"
onclick="magnifyFigure(&#39;habitual-ethics&#39;)">
<img class="img-button" src="{{ '/assets/images/Magnify_Large.svg' | relative_url }}" style="width:1.5ex">
</div>
<div id="habitual-ethics-caption" class="caption-frame">
<p>Figure: Launch event for Sylvie Delacroix’s book on habitual ethics
which covers the challenges of ethical thinking when we are devolving
decision making to habit.</p>
</div>
</div>
<p>But at the conference, the main I deas I wanted to get across was how
we might merge the data driven and the mechanistic approach by viewing
System 2 as a mechanistic-like model, which is a new way of feeding data
to the data-driven component.</p>
<p>This means that the relationship between System One and System Two
would be like a surrogate model or an emulator, similar to work seen in
Bayesian optimisation and other emulators. (Since then at Amazon the
team, led by Andrei Paleyes and Javier Gonzalez built a framework for
statistical emulation, <a href="https://emukit.github.io/">Emukit</a>,
that we also place at the core of one of the Masters courses I <a
href="https://mlatcl.github.io/mlphysical/">teach on in
Cambridge.</a>.</p>
<div class="figure">
<div id="bayes-opt-figure" class="figure-frame">
<div class="centered" style="">
<img class="" src="https://inverseprobability.com/slides/diagrams//ml/bayes-opt.png" width="40%" height="auto" align="center" style="background:none; border:none; box-shadow:none; display:block; margin-left:auto; margin-right:auto;vertical-align:middle">
</div>
</div>
<div id="bayes-opt-magnify" class="magnify"
onclick="magnifyFigure(&#39;bayes-opt&#39;)">
<img class="img-button" src="{{ '/assets/images/Magnify_Large.svg' | relative_url }}" style="width:1.5ex">
</div>
<div id="bayes-opt-caption" class="caption-frame">
<p>Figure: A representation of Bayesian optimisation, which is an
approach where a surrogage model (or emulation) is used as a proxy to
optimise rather than the slow to compute original system.</p>
</div>
</div>
<p>But key amoungst the surrogate, is the importance of “doubt”, the
ability of System One to know when it doesn’t know, and take a sensible
action. That was a theme I’d discussed in this blog post on blog post on
<a
href="http://inverseprobability.com/2015-11-09-artificial-stupidity">Artificial
Stupidity</a>.</p>
<p>The final slide hints at the workshop we had yesterday, on Causal
Digital Twins. Because what it says is that (roughly) our System Two
could be thought of as a (causal) digital twin of the real world, that
we use for planning. I think that relates very nicely to Bernhard’s
Konrad Lorenz quote thinking is acting in an imagined space.</p>
<div class="centered" style="">
<img class="" src="https://inverseprobability.com/slides/diagrams//systemOneAsSurrogate.png" width="80%" height="auto" align="center" style="background:none; border:none; box-shadow:none; display:block; margin-left:auto; margin-right:auto;vertical-align:middle">
</div>
<p>The theme for this section is caught by surprise, so what’s intesting
to me is how much of what I was proposing in this talk is work that we
continue to focus on today. However, if I was to be asked what caught me
by surprise, it’s the extent to which the data driven approach can be
extrapolated. I think the recent results aren’t a surprise given GPT-2,
but I think if you’d asked any of the people on that panel at the time
they wouldn’t have predicted the capabilities of GPT-2 with very little
“weak mechanism” in the model.</p>
<h2 id="thanks">Thanks!</h2>
<p>For more information on these subjects and more you might want to
check the following resources.</p>
<ul>
<li>twitter: <a href="https://twitter.com/lawrennd">@lawrennd</a></li>
<li>podcast: <a href="http://thetalkingmachines.com">The Talking
Machines</a></li>
<li>newspaper: <a
href="http://www.theguardian.com/profile/neil-lawrence">Guardian Profile
Page</a></li>
<li>blog: <a
href="http://inverseprobability.com/blog.html">http://inverseprobability.com</a></li>
</ul>
<h1 class="unnumbered" id="references">References</h1>
<div id="refs" class="references csl-bib-body hanging-indent"
role="doc-bibliography">
<div id="ref-Hilbert:information11" class="csl-entry"
role="doc-biblioentry">
Hilbert, M., López, P., 2011. The world’s technological capcity to
store, communicate and compute information. Science 332, 60–65. <a
href="https://doi.org/10.1126/science.1200970">https://doi.org/10.1126/science.1200970</a>
</div>
</div>