Research teams mine the intersection of neuroscience and AI

Stanford’s Wu Tsai Neurosciences Institute and Institute for Human-Centered Artificial Intelligence have awarded the second round of AI and Neuroscience Grants to five projects working at the intersection of these disciplines.

The links between neuroscience and artificial intelligence run deep: The very first artificial neural network models in the 1940s and 50s were both inspired by and intended to describe the biological neural networks in our brains. While the fields went their separate ways for decades, in the past fifteen years or so they’ve been increasingly cross-fertilizing. 

First launched in 2022, the collaborative AI and Neuroscience grant program aimed to foster exactly these kinds of interactions, starting with one-year, $125,000 awards to four teams studying stroke, Parkinson’s disease, speech decoding, and the fine-scale structure of links between neurons.

This time around, the program will fund five teams working on an even wider range of topics, including an effort to train an AI model based on how infants learn and another that will work to incorporate smell – and the emotional context that comes with it – into machine learning.

“This is an incredible moment for both neuroscience and AI research,” said Kang Shen, the Vincent V.C. Woo Director of the Wu Tsai Neurosciences Institute and the Frank Lee and Carol Hall Professor and professor of biology and of pathology in the Stanford schools of Humanities and Sciences and of Medicine. “We’re excited to take advantage of this moment and partner with HAI to support these projects, which we believe will drive both a deeper understanding of the brain and ethical advances in human-inspired AI.”

James Landay, the Denning Director of Stanford HAI, the Anand Rajaraman and Venky Harinarayan Professor and a professor of Computer Science at Stanford Engineering, said the projects aim to benefit both science and the world at large.  “At HAI, we’ve always taken a human-centered and interdisciplinary approach,” Landay said. “Working with Wu Tsai Neuro, we’re encouraging collaborative team science that will advance neuroscience, AI, and ultimately humanity.”

Funded projects

An AI-grounded framework for EEG, fMRI and joint EEG–fMRI depression biotypes

Cammie Rolle, Psychiatry and Behavioral Sciences
Ellen Kuhl, Mechanical Engineering, Catherine Holman Johnson Director of Stanford Bio-X

Recent research shows that fMRI brain scans can reveal different subtypes of depression, which suggests that psychiatrists might need to treat each subtype differently. Rolle and Kuhl will use AI methods to link fMRI results to another type of brain scan, electroencephalography (EEG), then evaluate whether EEG or a combination of methods might be used to refine these depression subtypes and guide better treatments.

How do infants become visual experts? Modeling infant experiences and brain data

Cameron Ellis, Psychology
Michael Frank, Psychology
Dan Yamins, Psychology, Wu Tsai Neuro

Human infants learn faster than AI systems, despite experiencing a small and biased fraction of the data that AI systems are trained on. Ellis, Frank, and Yamins will use fMRI brain scans and video taken from a baby’s point of view to study how infants learn so quickly. Specifically, they will build an AI system that is trained on both the brain data and video data of an infant to make the model learn like an infant does. In doing so, they will better understand how babies learn and also how AI systems can learn more efficiently.

Cracking the code of neuronal cell type identity

Marius Wernig, Pathology
Anshul Kundaje, Genetics and Computer Science

How a developing human brain produces all manner of different cell types – despite those cells being essentially genetically identical – remains one of the key mysteries of developmental biology. The matter is complicated by the complexity of the human genome. Wernig and Kundaje will draw on new developments in genetics and AI to build a foundation model of neuronal cell types to identify the molecular basis of neuronal cell type differentiation.

Building the neural atlas of psychotherapy: Mapping therapeutic dialogue to deep brain dynamics via large language models and topology

Corey Keller, Psychiatry and Behavioral Sciences
Manish Saggar, Psychiatry and Behavioral Sciences

Although psychotherapy is widely used to treat psychiatric disorders, little is known about how therapeutic conversations influence brain activity in real time. Keller and Saggar will record intracranial electrical activity from patients – who are also undergoing epilepsy monitoring – while they participate in psychotherapy sessions. Using large language models to characterize therapeutic dialogue and topological methods to map evolving brain states, they will investigate how different elements of therapeutic conversations relate to neural dynamics, with the long-term goal of building a neural atlas of psychotherapy to inform more effective treatments.

From molecules to memories: Aligning AI odor maps with human brain activity and personal experience

Todd Coleman, Bioengineering, Wu Tsai Neuro
Emily Fox, Statistics and Computer Science

Current machine learning algorithms excel at pattern recognition but struggle when it comes to capturing the subjective, context-dependent aspects of human intelligence. One way to address the struggle, Coleman and Fox argue, is smell. Smell is an underappreciated aspect of our daily experience, one that lies beyond the ken of the AI systems we are designing to work with us. The pair will use electromagnetic recordings from the brain and large language models to link brain activity to individuals’ subjective experiences of smell. Establishing a link between these experiences and neural encodings could reveal new principles for building AI systems that better understand human experiences.