#  Research 

 



##  Research Description 

 Our research areas are theoretical neuroscience and neuroscience-guided machine learning. We seek to uncover the algorithms of the brain and their implementation at the network and cellular levels.

 How can we infer what the brain computes from the large datasets of modern neuroscience? For that, we need a new “algorithmic” theory that bridges computation and its biological realization. Such a theory would systematically predict neural circuits that perform a given algorithm. It would predict a computational goal given a neural circuit. Despite our extensive knowledge of neuron physiology, there is not a commonly accepted algorithmic theory of neural computation.

 We have been working on an algorithmic theory for learning in the sensory domain. Sensory cortices learn from stimuli to build behaviorally relevant representations, with little or no supervision. Our theory starts by posing computational goals of unsupervised learning in the brain as mathematical optimization problems. Then, from these problems, we systematically derive algorithms and neural circuit implementations of these algorithms, linking computation to biological realization. The key contribution that made this program possible was our introduction of the so-called similarity-based cost functions. These cost functions contain a term that matches the similarity of outputs to the similarity of inputs. Online optimization of such similarity-based cost functions leads to algorithms implementable by neural networks with biologically plausible learning rules.

 So far, our approach achieved the following: 1) It provided new optimization formulations of common unsupervised learning tasks. 2) It answered how efficient self-organization for learning happens with local synaptic plasticity. 3) It provided new circuit motifs and mechanisms that could be in use in the brain. 4) It made experimental predictions about computations performed in specific circuits and provided computational interpretations of salient features of such circuits.

 The described theory is also a path to neuroscience-guided machine learning. The neural algorithms we uncovered, so far, solve unsupervised learning tasks such as linear dimensionality reduction, sparse and/or nonnegative feature extraction, blind nonnegative source separation, clustering and manifold learning. Many of these algorithms are on par in performance with state-of-the-art machine learning. We have found that respecting biological constraints, which may be naively viewed as a handicap, can instead facilitate the development of artificial neural networks by restricting the search space of possible algorithms.

 We also have broader interests in theoretical approaches to neuroscience. We study reinforcement learning of motor skills to infer how the brain solves this computationally hard problem at the network level. We figured out how stimulus selectivity could emerge in a network with random connectivity. We actively look for collaborations with experimentalists and projects motivated by new experimental results. This led us, for example, to publish a model of how the songbird brain might recover from lesions and a neuromechanical model of the fly larva’s crawling.