NeuroAI: A framework for deriving bio-plausible machine learning algorithms

by Mitya Chklovskii (Flatiron)

Friday Nov 11, 2022, 3:00 PM → 5:30 PM America/New_York

Ingrid Daubechies Auditorium/2-IDA (162 5th Avenue)

Time: 3-4 pm (Seminar) + 4-5:30 pm (Reception)

Location: 162 Fifth Ave, 2nd floor auditorium (Seminar) + 162 Rooftop or Promenade, depending on weather (Reception)

 

Abstract: Conventional artificial neural networks resemble computation in our brains only superficially. Successful machine learning algorithms like backpropagation violate fundamental biophysical observations, suggesting that our brains employ other algorithms to analyze high-dimensional datasets streamed from our sensory organs. We present a novel framework for deriving bio-plausible machine learning algorithms and neural networks. Our approach is normative, relying on the optimization of principled objective functions. Similarity-matching objective functions lead to neural networks relying exclusively on biologically plausible local learning rules and solving important unsupervised learning tasks such as dimensionality reduction, clustering, manifold learning, canonical correlation analysis etc. In addition to modeling biological networks, similarity-matching algorithms are competitive with conventional artificial neural networks.