Activities

About the Speaker: Dr. Jonathan D. Cohen is the Robert Bendheim and Lynn Thoman Professor of Neuroscience at Princeton University and was a founding co-director of the Princeton Neuroscience Institute. He has three decades of research experience in computational and mathematical modeling as well as empirical studies of human brain function and behavior, focusing on the neural mechanisms responsible for cognitive control and human intelligence, and how our growing understanding of these can be brought to bear in the design of computational architectures with more human-like cognitive capabilities.

Event Details: September 24th, 2026 - at 16:15 in Room 50/07

Title & Abstract: 

“Miller-Shepard’s Law" and the Curse of Generalization: A Unifying Information Theoretic Principle Underlying Capacity Constraints in Cognitive Function

The striking constraints of some human cognitive processes — including working memory and cognitive control — stand in stark contrast to the near limitless capability of others. While we can acquire and flexibly use vast amounts of information acquired over a lifetime, the amount we can process at any one time is often stiflingly limited — a limit famously referred to as the Magic Number 7 (+/- 2) by George Miller in his classic 1956 article. Here, we integrate ideas from information theory, cognitive science, and neuroscience to offer a unified account of why processing is often so limited. We show that there is a fundamental tradeoff between: generalization (how well existing representations can be leveraged process novel information), a characteristic strength of human cognition; and processing capacity (how many representations of distinct items can be processed at the same time), an equally characteristic cognitive weakness that has been observed in tasks ranging from perception, attention and working memory to multitasking capabilities and cognitive control. We demonstrate this fundamental tradeoff first in abstract information theoretic form, and then in a neurally motivated model that simulates performance in the classic cognitive tasks that motivated Miller to ponder the magicality of the number 7 (+/- 2). The generality of this principle also has profound implications for artificial systems, evidence for which will also be discussed.

About the Speaker: Dr. Viviane Clay is the Executive Director of the Thousand Brains Project (Numenta)

Event Details: August 24th, 2026 - 1pm sharp, 93/E44

Title & Abstract: 

“Thousand-Brains Systems: Sensorimotor Intelligence for Rapid, Robust Learning and Inference”

Current AI systems achieve impressive performance on many tasks, yet they lack core attributes of biological intelligence, including rapid, continual learning, representations grounded in sensorimotor interactions, and structured knowledge that enables efficient generalization. Neuroscience theory suggests that mammals evolved flexible intelligence through the replication of a semi-independent, sensorimotor module, a functional unit known as a cortical column. To address the disparity between biological and artificial intelligence, we are building a new type of machine intelligence grounded in the principles outlined in the Thousand Brains Theory of the neocortex. We have demonstrated various benefits of this novel approach over deep learning, including eight orders of magnitude fewer computations for training, no catastrophic forgetting, the ability to learn from very little data, and generalization outside the training distribution ( https://direct.mit.edu/neco/article/38/6/845/136222/Thousand-Brains-Systems-Sensorimotor-Intelligence). These capabilities are crucial in a variety of real-world sensorimotor applications, where data is scarce, robustness and interpretability are required, or compute is limited. I will present our open-source project ( https://thousandbrains.org/), its approach and benefits, as well as an overview of the neuroscience theory behind it.

About the Speaker:  Dr. H.S. (Steven) Scholte is a Professor of Psychology at the Faculty of Social and Behavioural Sciences, Programme group Brain and Cognition at the University of Amsterdam) 

Event Details: July 8th, 2026 - time and room t.b.a.

Title & Abstract: 

“Lunch Can Be Surprisingly Affordable: Neuroconnectionism as a Paradigm Shift in the Psychology of Vision”

The binding problem has organised vision research for forty years. If colour, motion, and form are extracted by separate modules, the brain must recombine them, and that recombination becomes the central problem perception has to solve at every moment. I will argue that neuroconnectionism does not solve this problem; it dissolves it, and that this dissolution is a worked example of what a paradigm shift does to a science. Binding does not survive translation across paradigms. Its status, whether central problem, technical challenge, or non-question, follows from each framework's commitments about how the visual world is represented.

I will discuss three claims. First, world structure is the key variable. The standard objection is that there is no free lunch: any code buying compositional generalisation must pay for it somewhere. That accounting assumes a system handling arbitrary feature combinations. Systems tuned to natural image statistics inherit composition from learned representational geometry rather than paying for it through an explicit binding operation, because the world supplies the structure the code would otherwise have to build. Lunch is affordable. Binding becomes a problem the system must solve only in a small set of cases, namely occlusion, clutter, and artificial displays, a ~1% problem rather than the ~99% the classical literature assumed. Second, specialisation is not segregation. Mixed-selectivity populations and a 307-patient lesion study show that selective deficits index high-gain participation rather than modular feature stores, so the inference from deficit to module fails. Third, binding for whom? The classical problem presupposes an internal observer who inspects bound representations. Neuroconnectionism replaces that question with a functional one: can downstream circuits drive appropriate behaviour from the population activity?

I will close by setting the frameworks side by side, on what each takes representation, composition, and attention to be and where each breaks. The contrast makes the central claim concrete: binding is not one problem with rival solutions but a different object under each set of representational commitments. Because the binding problem was so central to Good Old-Fashioned Vision, its dissolution is itself the signature of a paradigm shift.

About the Speaker:  Dr. Dirk Wulff is a Senior Research Scientist at the Center for Adaptive Rationality at the Max Planck Institute for Human Development, Berlin) 

Event Details: June 3, 2026 - at 16:30, Room: 50/E07 

Title & Abstract: 

“Clarifying the structure of psychology using LLMs”

Psychology has a conceptual clarity problem: thousands of constructs and measures with unclear boundaries, overlapping definitions, and inconsistent terminology, often referred to as jingle-jangle fallacies. Sorting out these relationships empirically would require prohibitively costly research designs. In this talk, I show how large language models can help map the relationships between psychological constructs, measures, and observable behavior at a scale that was previously impractical. I first present work showing that semantic embeddings from language models can predict empirical associations between survey items and scales, automatically detect jingle-jangle fallacies, and suggest more parsimonious taxonomies. I then present ongoing work that extends this approach beyond survey-based measures to behavioral tasks using task embeddings, vector representations that capture not only what a task asks but how people perform it. Together, these lines of work validate a generative approach that can help bring clarity and map out psychology's nomological net.

About the Speaker:  Randall C. O'Reilly is a Professor of Psychology and Computer Science at the Center for Neuroscience at the University of California, Davis (UC Davis) 

Event Details: May 27th, 2026 - at 16:30, Room: 50/E07 

Title & Abstract: 

“A simulated rat pauses to contemplate its future”

There is now overwhelming evidence that rats and other mammals make their decisions at the start of task trials, and spend relatively little time thereafter. This is consistent with the Rubicon theory (Heckhausen and Gollwitzer, 1987) thatpostulates two qualitatively distinct phases of mental life: goal selection vs. goal engaged. Among the many other implications of this theory, we procrastinate because once goal-enaged, we will be committed to finishing, or else suffer disappointment. The goal selection process is therefore especially conservative, and it is thus relatively difficult to actually cross the Rubicon into the goal engaged state. I have been developing a large-scale, systems-neuroscience computational model of the mechanisms that drive this and related goal-driven dynamics in the brain, which can shape learning and online processing in powerful ways. It requires the coordinated activity of many brain systems and neuromodulatory pathways to establish thisstrong future-oriented bias in learning and processing, and it is thus unlikely that these dynamics could simply emerge from generic error-driven learning mechanisms. Indeed,current deep neural networks powered by such mechanisms notably lack evidence of goal-driven, self-motivated behavior. Thus, it may be necessary to reverse-engineer the millions of years of evolution that shaped the mammalianbrain to understand how goal-driven learning works, and how it all-too-often breaks down in a wide range of mental disorders that plague humanity.