Interpretable AI as a New Subject

Jim talks with John Krakauer—professor of neurology and neuroscience, director of the Center for Study of Motor Learning and Brain Repair at Johns Hopkins, and external faculty at SFI—about his 2017 paper "Neuroscience Needs Behavior: Correcting a Reductionist Bias." They discuss defining behavior as ecologically valid goal-directed action within an animal's umwelt, behavioral decomposition being epistemically prior to neural investigation, bipedal running and Sherrington's spinalized cat experiments as illustrations of that decomposition, what a satisfying neural explanation should actually look like, emergence and neuroscientists' resistance to it, the concept of explanatory autonomy and the "wings don't fly, birds do" framing, downward causality and the traffic jam analogy, Sherrington's own epistemic humility about understanding thought, whether consciousness will eventually be explained the way life was or remain permanently fuzzy, the three traditions of studying the nervous system and their persistent tensions, the problem of double-dipping with coarse-grained behavioral language in neural data, "filler verbs" like "involves" and "underlies" that add surplus meaning to a correlation without doing extra explanatory work, everyday pseudo-explanations like dopamine for unhappiness and oxytocin for love, the identity fallacy, LLMs as scientific sparring partners and critical reviewers, Krakauer's vertigo at the current moment and the possibility of retiring if AI generates better intuitions, interpretable AI as a new subject for neuroscience and psychology, Jim's own artificial consciousness project building a rudimentary white-tailed deer, distinguishing consciousness from cognition and sentience, separating the machinery of consciousness from its contents, Nagel's "What Is It Like to Be a Bat?" and echolocation as conscious content, multiple realizability and its being pervasive and fatal to naive reductionism, the mereological fallacy and mirror neurons as ground zero for multiple fallacies, Marr's three levels and the direction of the scientific project from behavioral goal to algorithm to neural implementation, the bradykinesia paper finding that Parkinson's patients move slowly because they want to move more slowly, the C. elegans connectome and the limits of that knowledge, the Jonas and Kording microprocessor paper, and much more. Episode Transcript "Neuroscience Needs Behavior: Correcting a Reductionist Bias", by John Krakauer "What Is It Like to Be a Bat?", by Thomas Nagel "Why Don't We Move Faster?", by Pietro Mazzoni, Anna Hristova, and John Krakauer "Could a Neuroscientist Understand a Microprocessor?", by Eric Jonas and Konrad Kording John Krakauer is currently John C. Malone Professor, Professor of Neurology, Neuroscience, and Physical Medicine and Rehabilitation, and Director of the Brain, Learning, Animation, and Movement Lab at The Johns Hopkins University School of Medicine. He is also an External Professor at the Santa Fe Institute and Director of the Centre for Restorative Neurotechnology at The Champalimaud Centre for the Unknown. His areas of research interest include experimental and computational studies of motor control and motor learning, long-term skill learning and its relation to higher cognitive processes, prediction and mechanisms of motor recovery after stroke, new neuro-rehabilitation approaches including immersive XR gaming with generative AI, robotics and invasive CNS stimulation, and philosophy of mind. He is slowly working on a new book on the mind, intelligence, and AI for Princeton University Press.