Brain Inspired

Hakwan, Steve, and I discuss many issues around the scientific study of consciousness. Steve and Hakwan focus on higher order theories (HOTs) of consciousness, related to metacognition. So we discuss HOTs in particular and their relation to other approaches/theories, the idea of approaching consciousness as a computational problem to be tackled with computational modeling, we talk about the cultural, social, and career aspects of choosing to study something as elusive and controversial as consciousness, we talk about two of the models they're working on now to account for various properties of conscious experience, and, of course, the prospects of consciousness in AI. For more on metacognition and awareness, check out episode 73 with Megan Peters.

• Hakwan's lab: Consciousness and Metacognition Lab.
• Steve's lab: The MetaLab.
• Twitter: @hakwanlau; @smfleming.
• Hakwan's brief Aeon article: Is consciousness a battle between your beliefs and perceptions?
• Related papers
  • An Informal Internet Survey on the Current State of Consciousness Science.
  • Opportunities and challenges for a maturing science of consciousness.
  • What is consciousness, and could machines have it?"
  • Understanding the higher-order approach to consciousness.
  • Awareness as inference in a higher-order state space. (Steve's bayesian predictive generative model)
  • Consciousness, Metacognition, & Perceptual Reality Monitoring. (Hakwan's reality-monitoring model a la generative adversarial networks)

Timestamps

Brian and I discuss a range of topics related to his latest book, The Alignment Problem: Machine Learning and Human Values. The alignment problem asks how we can build AI that does what we want it to do, as opposed to building AI that will compromise our own values by accomplishing tasks that may be harmful or dangerous to us. Using some of the stories Brain relates in the book, we talk about:

• The history of machine learning and how we got this point;
• Some methods researches are creating to understand what's being represented in neural nets and how they generate their output;
• Some modern proposed solutions to the alignment problem, like programming the machines to learn our preferences so they can help achieve those preferences - an idea called inverse reinforcement learning;
• The thorny issue of accurately knowing our own values- if we get those wrong, will machines also get it wrong?

Links:

• Brian's website.
• Twitter: @brianchristian.
• The Alignment Problem: Machine Learning and Human Values.
• Related papers
  • Norbert Wiener from 1960: Some Moral and Technical Consequences of Automation.

Timestamps:

Omri, David and I discuss using recurrent neural network models (RNNs) to understand brains and brain function. Omri and David both use dynamical systems theory (DST) to describe how RNNs solve tasks, and to compare the dynamical stucture/landscape/skeleton of RNNs with real neural population recordings. We talk about how their thoughts have evolved since their 2103 Opening the Black Box paper, which began these lines of research and thinking. Some of the other topics we discuss:

• The idea of computation via dynamics, which sees computation as a process of evolving neural activity in a state space;
• Whether DST offers a description of mental function (that is, something beyond brain function, closer to the psychological level);
• The difference between classical approaches to modeling brains and the machine learning approach;
• The concept of universality - that the variety of artificial RNNs and natural RNNs (brains) adhere to some similar dynamical structure despite differences in the computations they perform;
• How learning is influenced by the dynamics in an ongoing and ever-changing manner, and how learning (a process) is distinct from optimization (a final trained state).
• David was on episode 5, for a more introductory episode on dynamics, RNNs, and brains.

• Barak Lab
• Twitter: @SussilloDavid
• The papers we discuss or mention:
  • Sussillo, D. & Barak, O. (2013). Opening the Black Box: Low-dimensional dynamics in high-dimensional recurrent neural networks.
  • Computation Through Neural Population Dynamics.
  • Implementing Inductive bias for different navigation tasks through diverse RNN attrractors.
  • Dynamics of random recurrent networks with correlated low-rank structure.
  • Quality of internal representation shapes learning performance in feedback neural networks.
  • Feigenbaum's universality constant original paper: Feigenbaum, M. J. (1976) "Universality in complex discrete dynamics", Los Alamos Theoretical Division Annual Report 1975-1976
• Talks
  • Universality and individuality in neural dynamics across large populations of recurrent networks.
  • World Wide Theoretical Neuroscience Seminar: Omri Barak, January 6, 2021

Timestamps:

K, Josh, and I were postdocs together in Jeff Schall's and Geoff Woodman's labs. K and Josh had backgrounds in psychology and were getting their first experience with neurophysiology, recording single neuron activity in awake behaving primates. This episode is a discussion surrounding their reflections and perspectives on neuroscience and psychology, given their backgrounds and experience (we reference episode 84 with György Buzsáki and David Poeppel). We also talk about their divergent paths - K stayed in academia and runs an EEG lab studying human decision-making and memory, and Josh left academia and has worked for three different pharmaceutical and tech companies. So this episode doesn't get into gritty science questions, but is a light discussion about the state of neuroscience, psychology, and AI, and reflections on academia and industry, life in lab, and plenty more.

• The Fukuda Lab.
• Josh's website.
• Twitter: @KeisukeFukuda4

Time stamps

It's generally agreed machine learning and AI provide neuroscience with tools for analysis and theoretical principles to test in brains, but there is less agreement about what neuroscience can provide AI. Should computer scientists and engineers care about how brains compute, or will it just slow them down, for example? Chris, Sam, and I discuss how neuroscience might contribute to AI moving forward, considering the past and present. This discussion also leads into related topics, like the role of prediction versus understanding, AGI, explainable AI, value alignment, the fundamental conundrum that humans specify the ultimate values of the tasks AI will solve, and more. Plus, a question from previous guest Andrew Saxe. Also, check out Sam's previous appearance on the podcast.

• Chris's lab: Human Information Processing lab.
• Sam's lab: Computational Cognitive Neuroscience Lab.
• Twitter: @gershbrain; @summerfieldlab
• Papers we discuss or mention or are related:
  • If deep learning is the answer, then what is the question?
  • Neuroscience-Inspired Artificial Intelligence.
  • Building Machines that Learn and Think Like People.

0:00 - Intro

Alison and I discuss her work to accelerate learning and thus improve AI by studying how children learn, as Alan Turing suggested in his famous 1950 paper. The ways children learn are via imitation, by learning abstract causal models, and active learning by implementing a high exploration/exploitation ratio. We also discuss child consciousness, psychedelics, the concept of life history, the role of grandparents and elders, and lots more.

• Alison's Website.
• Cognitive Development and Learning Lab.
• Twitter: @AlisonGopnik.
• Related papers:
  • Childhood as a solution to explore-exploit tensions.
  • The Aeon article about grandparents, children, and evolution: Vulnerable Yet Vital.
• Books:
  • The Gardener and the Carpenter: What the New Science of Child Development Tells Us About the Relationship Between Parents and Children.
  • The Scientist in the Crib: What Early Learning Tells Us About the Mind.
  • The Philosophical Baby: What Children's Minds Tell Us About Truth, Love, and the Meaning of Life.

Take-home points:

• Children learn by imitation, and not just unthinking imitation. They pay attention to and evaluate the intentions of others and judge whether a person seems to be a reliable source of information. That is, they learn by sophisticated socially-constrained imitation.
• Children build abstract causal models of the world. This allows them to simulate potential outcomes and test their actions against those simulations, accelerating learning.
• Children keep their foot on the exploration pedal, actively learning by exploring a wide spectrum of actions to determine what works. As we age, our exploratory cognition decreases, and we begin to exploit more what we've learned.

Timestamps

Dileep and I discuss his theoretical account of how the thalamus and cortex work together to implement visual inference. We talked previously about his Recursive Cortical Network (RCN) approach to visual inference, which is a probabilistic graph model that can solve hard problems like CAPTCHAs, and more recently we talked about using his RCNs with cloned units to account for cognitive maps related to the hippocampus. On this episode, we walk through how RCNs can map onto thalamo-cortical circuits so a given cortical column can signal whether it believes some concept or feature is present in the world, based on bottom-up incoming sensory evidence, top-down attention, and lateral related features. We also briefly compare this bio-RCN version with Randy O'Reilly's Deep Predictive Learning account of thalamo-cortical circuitry.

• Vicarious website - Dileeps AGI robotics company.
• Twitter: @dileeplearning
• The papers we discuss or mention:
  • A detailed mathematical theory of thalamic and cortical microcircuits based on inference in a generative vision model.
  • From CAPTCHA to Commonsense: How Brain Can Teach Us About Artificial Intelligence.
• Probabilistic graphical models.
• Hierarchical temporal memory.

Time Stamps:

0:00 - Intro

Russ and I discuss cognitive ontologies - the "parts" of the mind and their relations - as an ongoing dilemma of how to map onto each other what we know about brains and what we know about minds. We talk about whether we have the right ontology now, how he uses both top-down and data-driven approaches to analyze and refine current ontologies, and how all this has affected his own thinking about minds. We also discuss some of the current  meta-science issues and challenges in neuroscience  and AI, and Russ answers guest questions from Kendrick Kay and David Poeppel.

• Russ’s website.
• Poldrack Lab.
• Stanford Center For Reproducible Neuroscience.
• Twitter: @russpoldrack.
• Book:
  • The New Mind Readers: What Neuroimaging Can and Cannot Reveal about Our Thoughts.
• The papers we discuss or mention:
  • Atlases of cognition with large-scale human brain mapping.
  • Mapping Mental Function to Brain Structure: How Can Cognitive Neuroimaging Succeed?
  • From Brain Maps to Cognitive Ontologies: Informatics and the Search for Mental Structure.
  • Uncovering the structure of self-regulation through data-driven ontology discovery
• Talks:
  • Reproducibility: NeuroHackademy: Russell Poldrack - Reproducibility in fMRI: What is the problem?
  • Cognitive Ontology: Cognitive Ontologies, from Top to Bottom
  • A good series of talks about cognitive ontologies: Online Seminar Series: Problem of Cognitive Ontology.

Some take-home points:

• Our folk psychological cognitive ontology hasn't changed much since early Greek Philosophy, and especially since William James wrote about attention, consciousness, and so on.
• Using encoding models, we can predict brain responses pretty well based on what task a subject is performing or what "cognitive function" a subject is engaging, at least to a course approximation.
• Using a data-driven approach has potential to help determine mental structure, but important human decisions must still be made regarding how exactly to divide up the various "parts" of the mind.

Time points

Carsen and I discuss how she uses 2-photon calcium imaging data from over 10,000 neurons to understand the information processing of such large neural population activity. We talk about the tools she makes and uses to analyze the data, and the type of high-dimensional neural activity structure they found, which seems to allow efficient and robust information processing. We also talk about how these findings may help build better deep learning networks, and Carsen's thoughts on how to improve the diversity, inclusivity, and equality in neuroscience research labs. Guest question from Matt Smith.

• Stringer Lab.
• Twitter: @computingnature.
• The papers we discuss or mention:
  • High-dimensional geometry of population responses in visual cortex
  • Spontaneous behaviors drive multidimensional, brain-wide population activity.

Timestamps:

0:00 - Intro

Chris and I discuss his Spaun large scale model of the human brain (Semantic Pointer Architecture Unified Network), as detailed in his book How to Build a Brain. We talk about his philosophical approach, how Spaun compares to Randy O'Reilly's Leabra networks, the Applied Brain Research Chris co-founded, and I have guest questions from Brad Aimone, Steve Potter, and Randy O'Reilly.

• Chris's website.
• Applied Brain Research.
• The book: How to Build a Brain.
• Nengo (you can run Spaun).
• Paper summary of Spaun: A large-scale model of the functioning brain.

Some takeaways:

• Spaun is an embodied fully functional cognitive architecture with one eye for task instructions and an arm for responses.
• Chris uses elements from symbolic, connectionist, and dynamical systems approaches in cognitive science.
• The neural engineering framework (NEF) is how functions get instantiated in spiking neural networks.
• The semantic pointer architecture (SPA) is how representations are stored and transformed - i.e. the symbolic-like cognitive processing.

Time Points:

0:00 - Intro