Support the Podcast

Andrew and I discuss his work exploring how various facets of deep networks contribute to their function, i.e. deep network theory. We talk about what he’s learned by studying linear deep networks and asking how depth and initial weights affect learning dynamics, when replay is appropriate (and when it’s not), how semantics develop, and what it all might tell us about deep learning in brains.

Show notes:

• Visit Andrew's website.
• The papers we discuss or mention:
  • Are Efficient Deep Representations Learnable?
  • A theory of memory replay and generalization performance in neural networks.
  • A mathematical theory of semantic development in deep neural networks.
• A good talk: 
  • High-Dimensional Dynamics Of Generalization Errors.

A few recommended texts to dive deeper:

• Introduction To The Theory Of Neural Computation.
• Statistical Mechanics of Learning.
• Theoretical Neuroscience.

Support the Podcast

Jess and I discuss construction using graph neural networks. She makes AI agents that build structures to solve tasks in a simulated blocks and glue world using graph neural networks and deep reinforcement learning. We also discuss her work modeling mental simulation in humans and how it could be implemented in machines, and plenty more.

Show notes:

• Jess’s website.
• Follow her on twitter: @jhamrick
• The papers we discuss or mention:
  • Analogues of mental simulation and imagination in deeplearning.
  • Structured agents for physical construction.
  • Relational inductive biases, deep learning, and graph networks.
    • Build your own graph networks: Open source graph network library.

Kyle and I talk about his work modeling the basal ganglia and its circuitry to control whether we take an action and how we select among alternative actions. We also reflect on his experiences in academia, the larger picture of what it’s like in graduate school and after - at least in a computational neuroscience program - why he left, what he’s doing now, and how it all fits together.

Show notes:

• Kyle’s website.
• Follow him on twitter: @dunovank
• Examples of his work on basal ganglia and decision-making and control:
  • Believer-Skeptic Meets Actor-Critic: Rethinking the Role of Basal Ganglia Pathways during Decision-Making and Reinforcement Learning.
  • Reward-driven changes in striatal pathway competition shape evidence evaluation in decision-making.
  • Errors in Action Timing and Inhibition Facilitate Learning by Tuning Distinct Mechanisms in the Underlying Decision Process.
• Mark Humphries’ article on Medium: Academia is the Alternative Career Path.
• For fun, a bit about the “free will” experiments of Benjamin Libet. 

Support the Podcast

Phillip and I discuss his company Brainworks, which uses the latest neuroscience to build AI into its products. We talk about their first product, Ambient Biometrics, that measures vital signs using your smartphone's camera. We also dive into entrepreneurship in the AI startup world, ethical issues in AI and social media companies, his early days using neural networks at NASA, where he thinks this is all headed, and more.

Show notes:

• His company, Brainworks.
• Follow Phillip on twitter: @alvelda.
• Here's a talk he gave: Building Synthetic Brains.
• A guest post on Rodney Brooks's blog: Pondering the Empathy Gap.

Support the Podcast

Liz and I discuss her work on cognitive development, specially in infants, and what it can tell us about what makes human cognition different from other animals, what core cognitive abilities we’re born with, and how those abilities may form the foundation for much of our other cognitive abilities to develop. We also talk about natural language as the potential key faculty that synthesizes our early core abilities into the many higher cognitive functions that make us unique as a species, the potential for AI to capitalize on what we know about cognition in infants, plus plenty more.

Show notes:

• Visit Liz’s lab website.
• Related talks/lectures by Liz:
  • The Power and Limits of Artificial Intelligence.
  • A developmental perspective on brains, minds and machines.
• Visit the CCN conference website to learn more and see more talks.

In this second part of our conversation, (listen to the first part) David and I discuss his thoughts about current language and speech techniques in AI, his thoughts about the prospects of artificial general intelligence, the challenge of mapping the parts of linguistics onto the parts of neuroscience, the state of graduate training, and more.

• Visit David's lab website at NYU.
• He’s also a director at Max Planck Institute for Empirical Aesthetics.
• Follow him on twitter: @davidpoeppel.
• Some of the papers we discuss or mention (lots more on his website):
  • The cortical organization of speech processing.
  • The maps problem and the mapping problem: Two challenges for a cognitive neuroscience of speech and language.
• A good talk:
  • What Language Processing in the Brain Tells Us About the Structure of the Mind.
• Transformer model:
• How do Transformers Work in NLP? A Guide to the Latest State-of-the-Art Models.
• Attention Is All You Need.

Support the Podcast

David and I talk about his work to understand how sound waves floating in the air get transformed into meaningful concepts in your mind. He studies speech processing and production, language, music, and everything in between, approaching his work with steadfast principles to help frame what it means to understand something scientifically. We discuss many of the hurdles to understanding how our brains work and making real progress in science, plus a ton more.

• Visit David's lab website at NYU.
• He’s also a director at Max Planck Institute for Empirical Aesthetics.
• Follow him on twitter: @davidpoeppel.
• For a related episode (philosophically), you might re-visit my discussion with John Krakauer.
• Some of the papers we discuss or mention (lots more on his website):
  • The cortical organization of speech processing.
  • The maps problem and the mapping problem: Two challenges for a cognitive neuroscience of speech and language.
• A good talk:
  • What Language Processing in the Brain Tells Us About the Structure of the Mind.
• NLP Transformer model:
  • How do Transformers Work in NLP? A Guide to the Latest State-of-the-Art Models.
  • Attention Is All You Need.

Support the Podcast

Show notes:

Raia and I discuss her work at DeepMind figuring out how to build robots using deep reinforcement learning to do things like navigate cities and generalize intelligent behaviors across different tasks. We also talk about challenges specific for embodied AI (robots), how much of it takes inspiration from neuroscience, and lots more.

• Raia’s website.
• Follow her on Twitter: @RaiaHadsell
• Papers relevant to our discussion:
  • Learning to Navigate in Cities without a Map.
  • Overcoming catastrophic forgetting in neural networks.
  • Progressive neural networks.
• A few Talks:
  • Deep reinforcement learning in complex environments.
  • Progressive Nets & Transfer.
• The new Neuro-AI conference she's starting with Tony Zador and Blake Richards:
  • From Neuroscience to Artificially Intelligent Systems (NAISys) 

Talia and I discuss her work on how our visual system is organized topographically, and divides into three main categories: big inanimate things, small inanimate things, and animals. Her work is unique in that it focuses not on the classic hierarchical processing of vision (though she does that, too), but what kinds of things are represented along that hierarchy. She also uses deep networks to learn more about the visual system. We also talk about her keynote talk at the Cognitive Computational Neuroscience conference and plenty more.

Show notes:

• Talia’s lab website.
• Follow her on twitter: @talia_konkle.
• Check out the Cognitive Computational Neuroscience conference, where she'll give a keynote address.
• Papers we discuss/reference:
  • Early work on the tripartite organization. Tripartite Organization of the Ventral Stream by Animacy and Object Size.
  • A more recent update, with the texforms we discuss and comparision too deep learning CNN networks used to model the ventral visual stream. Mid-level visual features underlie the high-level categorical organization of the ventral stream.
• The article Talia references about an elegant solution to an old problem in computer science.

How does knowledge in the world get into our brains and integrated with the rest of our knowledge and memories? Anna and I talk about the complementary learning systems theory introduced in 1995 that posits a fast episodic hippopcampal learning system and a slower statistical cortical learning system. We then discuss her work that advances and adds missing pieces to the CLS framework, and explores how sleep and sleep cycles contribute to the process. We also discuss how her work might contribute to AI systems by using multiple types of memory buffers, a little about being a woman in science, and how it’s going with her brand new lab.

• Anna’s Penn Computational Cognitive Neuroscience Lab.
• Follow Anna on Twitter: @annaschapiro.
• Papers we discuss:
  • The original Complimentary Learning Systems paper: 
    • Complimentary Learning Systems Theory and Its Recent Update.
  • Anna’s work on CLS and Hippocampus:
    • The hippocampus is necessary for the consolidation of a task that does not require the hippocampus for initial learning.
    • Complementary learning systems within the hippocampus: a neural network modelling approach to reconciling episodic memory with statistical learning.
  • Examples of her work on sleep:
    • Active and effective replay: Systems consolidation reconsidered again.
    • Switching between internal and external modes: A multiscale learning principle.
    • Sleep Benefts Memory for Semantic Category Structure While Preserving Exemplar-Specifc Information.