Speaker Abstracts

---

Theories of learning, Imagination and Reasoning: of Mice and Machines

Surya Ganguli, Principal Investigator at Neural Dynamics and Computation Lab, Associate Professor, Stanford University

Abstract coming soon.

---

Computing with Neural Manifolds: Towards a Multi-Scale Understanding of Biological and Artificial Neural Networks

SueYeon Chung, Ph.D. Flatiron Institute, NYU

Recent breakthroughs in experimental neuroscience and machine learning have opened new frontiers in understanding the computational principles governing neural circuits and artificial neural networks (ANNs). Both biological and artificial systems exhibit an astonishing degree of orchestrated information processing capabilities across multiple scales – from the microscopic responses of individual neurons to the emergent macroscopic phenomena of cognition and task functions. At the mesoscopic scale, the structures of neuron population activities manifest themselves as neural representations. Neural computation can be viewed as a series of transformations of these representations through various processing stages of the brain. The primary focus of my lab’s research is to develop theories of neural representations that describe the principles of neural coding and, importantly, capture the complex structure of real data from both biological and artificial systems.

In this talk, I will present three related approaches that leverage techniques from statistical physics, machine learning, and geometry to study the multi-scale nature of neural computation. First, I will introduce new statistical mechanical theories that connect geometric structures that arise from neural responses (i.e., neural manifolds) to the efficiency of neural representations in implementing a task. Second, I will employ these theories to analyze how these representations evolve across scales, shaped by the properties of single neurons and the transformations across distinct brain regions. Finally, I will demonstrate how insights from the theories of neural representations can elucidate why certain ANN models better predict neural data, facilitating model comparison and selection

---

Using AI to Measure and Model Neural Dynamics & Behavior

Mackenzie Mathis, Ph.D. EPFL

Advances in AI are transforming how we study brain and behavior. This talk will cover my lab’s development of methods for measuring nonlinear neural dynamics and behavior with generalized contrastive learning, pose estimation, and multimodal vision-language models. Lastly, I will touch on how we use these tools to extract cortical computations learning during skill learning in mice.

---

Learning Compositional Models of the World

Yilun Du, Ph.D. Harvard University

A major bottleneck towards constructing intelligent embodied agents is the lack a of available data for all the settings the agent might find itself in. I’ll illustrate how we can construct systems that can operate well in such scenarios by building a model of the world and by then using inference on the model as a way to solve new tasks. I’ll further illustrate how we can build models compositionality, so that models can generalize in areas we do not have data in. I illustrate a set of results using such an approach across perception, reasoning, and decision making.

---

Sequence Prediction through Local Learning

João Sacramento, Ph.D. Google

Some autoregressive models exhibit in-context learning capabilities: being able to learn as an input sequence is processed, without undergoing any parameter changes, and without being explicitly trained to do so. To try to understand this phenomenon, we analyze neural sequence models trained on synthetic tasks. This analysis explains a number of key findings on in-context learning observed in neural language models, and suggests approaching sequence prediction problems by stacking local learners: layers whose neural dynamics are derived from a local learning objective. We study this approach in language modeling at the 1 billion parameter scale, and find that it performs competitively when compared to the state-of-the-art. We discuss dynamic test-time compute allocation as one possible advantage of the local learning approach.

---

Summary statistics of learning link changing neural representations to behavior

Cengiz Pehlevan, Ph.D. Harvard University

How can we make sense of large-scale recordings of neural activity across learning? Theories of neural network learning with their origins in statistical physics offer a potential answer: for a given task, there are often a small set of summary statistics that are sufficient to predict performance as the network learns. I will review recent advances in how summary statistics can be used to build theoretical understanding of neural network learning. I will then argue for how this perspective can inform the analysis of neural data, enabling better understanding of learning in biological and artificial neural networks.

---

Causality: Why Most Claims to Causality are Bogus and How to Sometimes get is Nonetheless

Konrad Körding, Ph.D. NeuroMatch & University of Pennsylvania

I will review the state of mechanistic claims in neuroscience. I will contribute to an intuition of when causal inference is possible and when it is not. In discussing this, I will highlight which aspects of the world make causal thinking so useful. And which aspects of the world make people believe that correlation and causation is the same. I will also briefly comment on the foundation model/ Causality link.

---

Illuminating Synaptic Learning

Karel Svoboda, Ph.D. Allen Institute for Neural Dynamics

How do synapses in the middle of the brain know how to adjust their weight to advance a behavioral goal? This is referred to as the synaptic ‘credit assignment problem’. A large variety of synaptic learning rules have been proposed, mainly in the context of artificial neural networks. The most powerful learning rules (e.g. back-propagation of error) are thought to be biologically implausible, whereas the widely studied biological learning rules (Hebbian) are insufficient for goal-directed learning. I will describe ongoing work, both experimental and theoretical, focused on understanding synaptic learning rules in the cortex.

---

Mixed-modal Language Modeling: Chameleon, Transfusion, and Mixture of Transformers

Luke Zettlemoyer, Ph.D. University of Washington & Meta

Existing multimodal models typically have custom architectures designed for specific modalities (image->text, text->image, text only, etc). In this talk, I will present our recent work on a series of early fusion mixed-modal models trained on arbitrary mixed sequences of images and text. I will discuss and contrast two models architectures, Chameleon and Transfusion, that make very different assumptions about how to model mixed-modal data, and argue for moving from a tokenize-everything approach to newer models that are hybrids of autoregressive transformers and diffusion. I will also cover recent efforts to better understand how to more stably train such models at scale without excessive modality competition, using a mixture of transformers technique. Together, these advances lay a possible foundation for universal models that can understand and generate data in any modality, and I will also sketch some of the steps that we still need to focus on to reach this goal.

---

What Came First, the Sum or the Parts? Emergent Compositionality in Neural Networks

Ellie Pavlick, Ph.D. Brown University & Google Deepmind

Decades of research in cognitive science and AI have focused on compositionality as a hallmark property of the human mind. This focus can seem to frame the question as though we must classify systems as either compositional or idiomatic, cognitive or associative. In this talk, I describe a set of related but different empirical studies of how neural networks achieve, or fail to achieve, compositional behavior. I argue that these findings point to a middle ground in which traditional “symbolic” compositionality can be seen as a special case which is emergent—but nonetheless qualitatively different—from a more general associative mechanism characteristic of neural networks.

---

Offline Cortical Reactivations of Recent Experiences

Mark Andermann, Ph.D. Beth Isreal Deaconess Medical Center & Harvard Medical School

In this talk, I will describe our recent experimental work that relates to two concepts in natural and artificial intelligence. The first concept involves offline cortical reactivations of recent experiences as a mechanism for efficiently driving learning from sparse samples. The second concept relates to biologically plausible alternatives to backpropagation – “target learning” or “prospective configuration” – in which the brain first infers the optimal target activity pattern associated with a stimulus and only then instructs weight modifications to gradually achieve this pattern during learning. To investigate these ideas, we imaged calcium activity of thousands of excitatory neurons in mouse visual cortex. Presentation of a stimulus resulted in transient, stimulus-specific reactivations during the following minute. Surprisingly, reactivations that occurred early in a session systemically differed from preceding activity patterns evoked by the stimulus. These reactivations were more similar to future patterns evoked by the stimulus, thereby predicting within-day and across-day representational drift. The rate and content of these reactivations was sufficient to accurately predict the dynamics of future changes in stimulus responses and, surprisingly, the decreasing similarity of patterns of responses to distinct stimuli. These findings hint at a role for offline reactivations of recent experiences in driving optimal neural representations of the sensory world.

---

Model Interpretability: from Illusions to Opportunities

Asma Ghandeharioun, Ph.D. Google Deepmind

• While the capabilities of today’s large language models (LLMs) are reaching—and even surpassing— what was once thought impossible, concerns remain regarding their misalignment, such as generating misinformation or harmful text, which continues to be an open area of research. Understanding LLMs’ internal representations can help explain their behavior, verify their alignment with human values, and mitigate instances where they produce errors. In this talk, I begin by challenging common misconceptions about the connections between LLMs’ hidden representations and their downstream behavior, highlighting several “interpretability illusions.”
• Next, I introduce Patchscopes, a framework we developed that leverages the model itself to explain its internal representations in natural language. I’ll show how it can be used to answer a wide range of questions about an LLM’s computation. Beyond unifying prior inspection techniques, Patchscopes opens up new possibilities, such as using a more capable model to explain the representations of a smaller model. I show how patchscope can be used as a tool for inspection, discovery, and even error correction. Some examples include fixing multihop reasoning errors, the interaction between user personas and latent misalignment, and understanding why different classes of contextualization errors happen.
• I hope by the end of this talk, the audience shares my excitement in appreciating the beauty of the internal mechanisms of AI systems, understands the nuances of model interpretability and why some observations might lead to illusions, and takes away Patchscope, a powerful tool for qualitative analysis of how and why LLMs work and fail in different scenarios.

---

An Analytic Theory of Creativity in Convolutional Diffusion Models

Mason Kamb, Stanford University

We obtain an analytic, interpretable and predictive theory of creativity in convolutional diffusion models. Indeed, score-based diffusion models can generate highly creative images that lie far from their training data. But optimal score-matching theory suggests that these models should only be able to produce memorized training examples. To reconcile this theory-experiment gap, we identify two simple inductive biases, locality and equivariance, that: (1) induce a form of combinatorial creativity by preventing optimal score-matching; (2) result in a fully analytic, completely mechanistically interpretable, equivariant local score (ELS) machine that, (3) without any training can quantitatively predict the outputs of trained full convolutional diffusion models (like ResNets and UNets) with high accuracy (median r2 of 0.90, 0.91, 0.94 on CIFAR10, FashionMNIST, and MNIST). Our ELS machine reveals a locally consistent patch mosaic model of creativity, in which diffusion models create exponentially many novel images by mixing and matching different local training set patches in different image locations. Our theory partially predicts the outputs of pre-trained self-attention enabled UNets (median r2 ∼ 0.75 on CIFAR10), revealing an intriguing role for attention in carving out semantic coherence from local patch mosaics. Our theory also reveals for the first time the mechanism behind common spatial inconsistencies in diffusion-generated images, theoretically predicting common defects such as incorrect numbers of limbs and digits as a consequence of excessive spatial locality in generation.

---

Can Your Neurons Hear the Shape of an Object?

Mozes Jacobs, Harvard University

Traveling waves of neural activity are widely observed in the brain, but their precise computational function remains unclear. One prominent hypothesis is that they enable the transfer and integration of spatial information across neural populations. However, few computational models have explored how traveling waves might be harnessed to perform such integrative processing. Drawing inspiration from the famous “Can one hear the shape of a drum?” problem — which highlights how normal modes of wave dynamics encode geometric information — we investigate whether similar principles can be leveraged in artificial neural networks. Specifically, we introduce convolutional recurrent neural networks that learn to produce traveling waves in their hidden states in response to visual stimuli, enabling spatial integration. By then treating these wave-like activation sequences as visual representations themselves, we obtain a powerful representational space that outperforms local feed-forward networks on tasks requiring global spatial context. In particular, we observe that traveling waves effectively expand the receptive field of locally connected neurons, supporting long-range encoding and communication of information. We demonstrate that models equipped with this mechanism solve visual semantic segmentation tasks demanding global integration, significantly outperforming local feed-forward models and rivaling non-local U-Net models with fewer parameters. As a first step toward traveling-wave-based communication and visual representation in artificial networks, our findings suggest wave-dynamics may provide efficiency and training stability benefits, while simultaneously offering a new framework for connecting models to biological recordings of neural activity.

---

Broadly-projecting Mesolimbic Dopamine Neurons Implement a Distributional Critic Across the Striatum

Sara Matias, Ph.D. Harvard University

Animal behavior is controlled through the coordinated action of multiple learning systems in the brain. One of these systems, the basal ganglia, instantiates a reinforcement learning (RL) algorithm in which dopamine (DA) neurons transmit reward prediction error (RPE) signals—the difference between actual and expected rewards—to enable value learning via cortico-striatal plasticity. Recent studies have highlighted two novel aspects: first, that RPE signals from midbrain DA neurons can encode entire reward distributions through a distributional RL algorithm that mirrors cutting-edge machine learning approaches, and second, that dopamine axons projecting to different regions of the striatum exhibit functional heterogeneity, indicating that not all DA neurons encode RPE. To examine the functional and anatomical organization of RPE and non-RPE dopamine signals, we conducted multi-fiber photometry recordings of dopamine axonal activity across the entire striatum. We observed that while RPE signals are present throughout the striatum in a reward-based task, aversive signals are heterogeneous. For example, DA in the dorsomedial striatum is activated by airpuffs, while that in the dorsolateral striatum conveys a brief biphasic response. However, fiber photometry recordings cannot disentangle whether the recorded signals are generated from a uniform population of dopamine axons or if, on the contrary, functionally heterogeneous axons intermingle in any particular striatal area. To overcome this limitation, we performed projection-identified electrophysiological recordings from midbrain DA neurons, to investigate if all dopamine neurons, projecting to all striatal regions, encode the reward distribution. We found that pure RPE-encoding DA neurons project to the lateral nucleus accumbens shell (lAcbSh), and broadly across the striatum. Moreover, lAcbSh- and broadly-projecting DA neurons show structured RPE heterogeneity consistent with distributional RL predictions for a quantile-like population code. Our findings suggest that dopamine-based RL is organized through a “distributional critic” architecture that is superimposed on other outcome-specific information, supporting continuous, reward-informed, behavioral control.

---

Excitatory-Inhibitory Dynamics in Adaptive Decision-Making

Veronica Chelu, McGill University

Neural circuits rely on excitatory–inhibitory (E/I) interactions to support adaptive learning and decision-making [5]. Here, we investigate the computational role of the E/I balance in shaping neural circuits for decision-making and continual reinforcement learning (RL). First, using a mean-field E/I model of two-choice decision-making [2], we show that excitatory recurrence enhances responsiveness by amplifying signals and driving circuits toward near-critical states, while inhibition stabilizes or refines decision-selective activity. Next, we integrate a bio-inspired E/I mechanism in an Actor-Critic[3] agent to examine how a well-tuned E/I balance supports adaptive learning in environments with shifting contingencies. By modulating inhibitory feedback, we show that E/I interactions regulate the speed-accuracy trade-off, enabling flexible control over adaptation rate. Our results suggest that E/I tuning is essential for maintaining an optimal balance between flexibility and stability, thus supporting efficient continual learning. We then compare two E/I recurrent neural network (RNN) approaches in a continual RL setting for decision-making tasks. A naive partitioning of excitatory and inhibitory populations [1] proves insufficient for dynamic tasks, resulting in poor performance. In contrast, Dale’s Artificial Neural Networks (DANNs)[4]—which incorporate biologically realistic E/I constraints—demonstrate robust learning and reliable adaptation across varying task demands. Finally, we discuss these findings in the context of neuromodulation, highlighting how E/I balance can serve as a fundamental mechanism for regulating cognitive function. Overall, our work underscores the computational importance of E/I interactions in shaping neural circuit dynamics, enabling adaptive decision-making and continual reinforcement learning in biologically inspired models.

---

Emergent Cognitive & Neural Alignment Between Biological and Artificial Systems

George Alvarez, Ph.D. Harvard University

Abstract coming soon.

---

What do AI chatbots think about us? Implications for user transparency and control

Fernanda Viegas, Ph.D. Insight and Interaction Lab & Harvard University

When you interact with a chatbot, what does it “think” about you? Does it care about your age, level of education, or socioeconomic status? It turns out the answer is yes. Recent work in AI interpretability is beginning to provide intriguing answers about how implicit social cognition helps shape chatbot behavior. In this talk I’ll describe an interface that translates the complex geometry of AI language models into a simple, understandable dashboard. This dashboard allows end users to see—and also to control—how a chatbot perceives them. I argue that this type of transparency is an important step in helping people work with AI more effectively, safely, and enjoyably.