David and I discuss the latest efforts he and his Elemental Cognition team have made to create machines that can understand stories the way humans can and do. The long term vision is to create what David calls “thought partners”, which are virtual assistants that can learn and synthesize a massive amount of information for us when we need that information for whatever project we’re working on. We also discuss the nature of understanding, language, the role of the biological sciences for AI, and more.
Rodrigo and I discuss concept cells and his latest book, NeuroScience Fiction. The book is a whirlwind of many of the big questions in neuroscience, each one framed by of one of Rodrigo’s favorite science fiction films and buttressed by tons of history, literature, and philosophy. We discuss a few of the topics in the book, like AI, identity, free will, consciousness, and immortality, and we keep returning to concept cells and the role of abstraction in human cognition.
In this second part of my conversion with Paul, we continue our discussion about how to understand brains as feedback control mechanisms – controlling our internal state and extending that control into the world – and how Paul thinks the key to understanding intelligence is to trace our evolutionary past through phylogenetic refinement.
In this first part of our conversation, Paul and I discuss his approach to understanding how the brain (and intelligence) works. Namely, he believes we are fundamentally action and movement oriented – all of our behavior and cognition is based on controlling ourselves and our environment through feedback control mechanisms, and basically all neural activity should be understood through that lens. This contrasts with the view that we serially perceive the environment, make internal representations of what we perceive, do some cognition on those representations, and transform that cognition into decisions about how to move. From that premise, Paul also believes the best (and perhaps only) way to understand our current brains is by tracing out the evolutionary steps that took us from our single celled first organisms all the way to us – a process he calls phylogenetic refinement.
Thomas and I discuss the role of recurrence in visual cognition: how brains somehow excel with so few “layers” compared to deep nets, how feedback recurrence can underlie visual reasoning, how LSTM gate-like processing could explain the function of canonical cortical microcircuits, the current limitations of deep learning networks like adversarial examples, and a bit of history in modeling our hierarchical visual system, including his work with the HMAX model and interacting with the deep learning folks as convolutional neural networks were being developed.
