Clustering of Neural Activity: A Design Principle for Population Codes

Princeton University, Princeton NJ

Investigators

Abstract

In virtually every part of the brain, information about the sensory environment, internal body states, or intended movements is encoded by more than one neuron. This was apparent as early as the nineteenth century from the extensive interconnectivity of nearby neurons and continues to be apparent from numerous measurements of the tuning curves and correlation of nearby neurons. Despite its fundamental importance, population neural codes are poorly understood. In this project the PI will test a recently developed principle for population code - namely, that neural activity patterns should always be organized into a discrete set of clusters. This organization is appealing, because clusters exhibit error correction, encode qualitatively different stimulus features than their constituent neurons, and can be learned in an unsupervised fashion by downstream neural circuits. Together, these properties enable a powerful form of hierarchical detection of complex stimulus features. The PI aims at developing a principle of population coding that applies widely across the brain. Furthermore, his hypothesis is connected to a model of hierarchical feature detection, which may be operating across all the ascending pathways in the neocortex. In addition, the neocortex's modular structure lends itself to machine learning algorithms. In addition,the project will involve the development and dissemination of software to perform fits of maximum entropy and hidden Markov models to neural data, which could help spur on the research programs of many labs that use these methods to analyze neural populations. This proposal will combine large-scale neural recording methods from the retina and cortex with state-of-the-art theoretical analyses to study collective phenomena in population neural codes. The PI has an extensive track record in multi-electrode recording from the vertebrate retina, along with applying maximum entropy models to analyze states of network activity. In addition, the PI has recently started a collaboration with Prof. David Tank to record from cortical populations using two-photon calcium fluorescence imaging. The project's aims are: 1. Study how cluster codes can simultaneously encode categorical and continuous stimulus information; if true, this idea would constitute a new example of multiplexed coding in neural populations 2. Study how biologically plausible neural networks can learn and readout clusters; if true, then this makes possible hierarchical cluster processing in the neocortex and will lead us to formulate specific experimental tests to see if layer 4 of V1 is learning LGN clusters. 3. Test whether cluster codes are present in central brain areas, like the visual cortex and hippocampus; this study may generalize his results beyond the retina and help establish a new design principle for population codes. This project is being jointly supported by the Physics of Living Systems program in the Division of Physics and the Modulation Program in the Division of Integrative Organismal Systems. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.