Research Themes

Neural mechanisms of relational long term memory formation

Our knowledge of everyday concepts and categories is incredibly relationally rich. We know that cats chase mice, have fur, and are animals, and understand the meanings of concepts that themselves capture relational structures, like communicate or desire. How do we infer and integrate relational structure from noisy, sensory experiences to build such knowledge? Understanding how we do so promises to shed light on age-old questions of the relationship between concepts, the structure of our minds, and the observable world.

In our approach to this question, we tackle a longstanding research gap in cognitive neuroscience at the juncture of episodic memory, relational inference, and semantic memory. Using learning paradigms followed by later neuroimaging of newly formed knowledge, we seek to elucidate the role of specialized episodic memory circuits in initially encoding relational structures and the pathways for eventually updating semantic memory areas.

  • Leshinskaya, A., Nguyen, M.A., & Ranganath, C. (2023). Integration of event experiences to build relational memory in the human brain. Cerebral Cortex. pdf

  • Leshinskaya, A., & Thompson-Schill, S.L. (2020). Transformation of event representations along middle temporal gyrus. Cerebral Cortex. 30(5), 3148–3166. pdf

  • Leshinskaya, A., Bajaj, M. & Thompson-Schill, S.L. (2023). Novel objects with causal schemas elicit selective responses in tool- and hand-selective lateraloccipito-temporal cortex. Cerebral Cortex, 33 (9), 5557-5573. pdf

Relational combination in neural networks

Relational knowledge relies on compositionality: the ability to meaningfully and flexibly combine smaller elements into structured combinations. It remains unsolved how a neural network can achieve this feat. In human brains, relational memory research has investigated how we learn that two elements are related, but less so how they are related. Yet we readily know more than just that cats and mice go together; we know that cats chase mice but fear dogs. We seek to test schemes for encoding relational type as developed in artificial neural networks, using fMRI in the human brain.

State of the art deep neural networks — especially large language models — are exhibiting unprecedented success in relational compositionality, but in ways we do not fully understand. With their greater visibility than human brains, they offer a unique opportunity to reverse-engineer a solution to one of the longstanding questions in cognitive science. We use the tools of mechanistic interpretability to answer these questions.

  • McCoy, M.B., & Leshinskaya, A.,. Relational composition during attribute retrieval in GPT is not purely linear. pdf

Moral reasoning as a test case for human-AI alignment

Moral reasoning draws on many of the most advanced aspects of human cognition, particularly relational reasoning and analogy. We seek to develop computational cognitive models that can shed light both on human moral reasoning and the moral computations emergent in complex AI models, such as LLMs. With tools for a rigorous computational/mechanistic description of moral reasoning, we can better measure, steer, and align the moral cognition of rapidly developing AI agents.

  • Leshinskaya, 2024. (blog post). Morally Guided Action Reasoning in Humans and Large Language Models: Alignment Beyond Reward.

  • Leshinskaya, A. & Chakroff, A. (2023). Value as semantics: representations of human moral and hedonic value in large language models. Advances in Neural Information Processing Systems (37), AI meets moral philosophy and moral psychology workshop. PDF