Ben Sorscher
Ben Sorscher is an AI researcher whose work connects theory for deep learning with neural systems in brains and artificial networks. He studied as an undergraduate at Harvard from 2014 to 2018, then coauthored an analytic account of grid cell patterns in 2019; in 2021, task-driven neural networks helped explain varied responses in the medial entorhinal cortex. A 2022 data pruning paper reported scaling better than a power law and received a NeurIPS Outstanding Paper Award; a separate 2022 paper studied learning with constrained weight distributions. In 2024, he coauthored a PNAS perspective on representation geometry and generalization, and, in October 2024, a Nature study that found stable grid cell firing patterns after one exposure to novel or altered environments. He is a PhD student at Stanford University whose research connects computational neuroscience with the theoretical foundations of deep learning.
Timeline
-
2014–2018
Undergraduate studies at Harvard
From 2014 to 2018, he studies as an undergraduate at Harvard University.
-
2019
A theory of grid cell patterns
In 2019, he and his coauthors propose an analytic theory that describes learning in neural networks trained on navigation tasks as a system that forms grid cell patterns.
-
2021
Task-driven entorhinal models
In 2021, he coauthors a study that uses task-driven neural networks to explain heterogeneous responses in the medial entorhinal cortex.
-
2022
Data pruning earns an award
In 2022, he coauthors a paper showing that data pruning lets ResNets trained on CIFAR-10, SVHN, and ImageNet scale better than a power law. NeurIPS gives the paper an Outstanding Paper Award.
-
2022
Learning with weight constraints
In 2022, he coauthors a statistical mechanical theory for learning in neural networks when weight distributions are constrained. The paper derives a perceptron’s memory capacity under those constraints and shows that this learning method outperforms methods with no constraint or only sign constraints.
-
2024
Representations and generalization
In 2024, he coauthors a PNAS perspective on generalization in biological and artificial deep neural networks. The paper studies how geometric properties of neural representations relate to generalization.
-
October 2024
One-shot entorhinal maps
In October 2024, he and John Wen contribute equally to a Nature study that records more than 15,000 grid cells in mice. It reports stable grid cell firing patterns in novel and altered environments after a single exposure.
-
Since 2018
A Stanford PhD student
Since 2018, he has been a PhD student at Stanford University, where his research connects computational neuroscience with the theoretical foundations of deep learning.
Sources and credits
Photo: We have no published photo of Ben Sorscher yet. Send us a free-licensed photo.
- Ben Sorscher (Stanford University) — Lacuna
- A unified theory for the origin of grid cells through the lens of pattern formation
- Explaining heterogeneity in medial entorhinal cortex with task-driven neural networks
- Beyond neural scaling laws: beating power law scaling via data pruning
- NeurIPS 2022 Awards
- A theory of weight distribution-constrained learning
- Representations and generalization in artificial and brain neural networks
- One-shot entorhinal maps enable flexible navigation in novel environments