Study of Theoretical and Computational Neuroscience
Statistical Mechanics of Deep Learning
Talk given at 2022 Foundations of Machine Learning and Its Applications for Scientific Discovery in Physical and Biological Systems Workshop. Hosted by Rice U. NSF funded.
Featured Work
Featured Program
Haim Sompolinsky is the Director of Harvard’s Swartz Program in Theoretical and Computational Neuroscience.
Featured Research
We can learn novel concepts from just one or a few sensory experiences. We demonstrate that concepts defined by tightly circumscribed manifolds in the neural firing-rate space of higher-order sensory areas can be learned from few examples when single plastic downstream readout neurons learn to discriminate using a simple plasticity rule.
Featured Project
We are seeking to advance the theoretical understanding of neural processing of structured knowledge (context, maps, relational reasoning), which is at the heart of high cognitive functions, using architectures and dynamical constraints (gating and binding) imposed by biology.
Latest News
- Haim’s interview in YediotHaim’s interview upon being awarded The Brain Prize (article in Hebrew)…
- Sompolinsky’s Computational Neuroscience CourseHaim is teaching Computational Neuroscience (Neuro/Physics/MCB 231) spring semester 2024…
- Haim Sompolinsky named Kempner Institute Associate FacultyThe Kempner Institute for the Study of Natural and Artificial Intelligence at Harvard announced the appointment of Haim Sompolinsky as one of the Institute’s inaugural cohort of associate faculty. …
- Grad Student and Post Doc OpeningsSee the Positions page for details….
- 2022 Gruber Foundation Neuroscience Prizehttps://gruber.yale.edu/haim-sompolinsky…
Featured Publications
- Optimal quadratic binding for relational reasoning in vector symbolic neural architecturesNaoki Hiratani, Haim Sompolinsky Neural Computation 35 (2), 105-155…
- Associative memory of structured knowledgeJulia Steinberg, Haim Sompolinsky Scientific Reports 12 (1), 21808…
- Minimum perturbation theory of deep perceptual learningHaozhe Shan, Haim Sompolinsky. Physical Review E 106 (6), 064406…
- Globally gated deep linear networksQianyi Li, Haim Sompolinsky arXiv:2210.1744…
- Neural representational geometry underlies few-shot concept learningBen Sorscher, Surya Ganguli, Haim Sompolinsky. Proc Natl Acad Sci U S A. 2022 Oct 25;119(43):e2200800119…
- Statistical Mechanics of Deep Linear Neural Networks: The Backpropagating Kernel RenormalizationQianyi Li, Haim Sompolinsky Phys. Rev. X, Volume 11, Issue 3, Pages 031059 (2021)…