Towards Localizing Computations in Distributed Cortical Networks
Chettih, Selmaan N.
MetadataShow full item record
CitationChettih, Selmaan N. 2019. Towards Localizing Computations in Distributed Cortical Networks. Doctoral dissertation, Harvard University, Graduate School of Arts & Sciences.
AbstractNeurons in mammalian neocortex form a spatially distributed network, where distant neurons interact on short temporal scales. Determining the extent to which the computational functions of this network are localized in distinct subsets is a major aim of systems neuroscience. In this work we present two complementary approaches towards meeting this aim.
The first approach utilizes a novel navigation-based decision making task for mice in virtual reality, designed to dissociate sensory stimuli, motor actions, and the animal’s knowledge of rules relating stimuli and actions to reward outcomes. We collected a dataset with neural activity from hundreds of thousands of neurons sampled densely and evenly across the dorsal posterior region of cortex, including visual, parietal, and retrosplenial areas. This permits quantitative comparison of the strength and population structure of neural encoding, with fine spatial resolution. Ongoing analyses indicate encoding properties form smooth and overlapping spatial gradients, with high quantitative precision.
The computational function of a neural circuit has a complex relationship with representations encoded in its activity. Thus the second approach uses single-neuron perturbations to estimate the functional effect of one neuron’s activity on another’s. By relating this influence to the encoding properties of each neuron, we infer a computational function of a neuron’s activity. We develop an all-optical implementation of this idea – influence mapping – which uses two-photon optogenetics to trigger action potentials in a targeted neuron and calcium imaging to measure the effect in neighboring neurons. In L2/3 of visual cortex, this method revealed a suppressive effect of excitatory neurons on their neighbors, with a center-surround profile over anatomical space. A neuron’s influence on its neighbor also depended on their similarity in activity. Neurons with similar representations of visual stimuli specifically suppressed each other’s activity, and reduced the gain of population responses to their preferred visual stimuli by ~2%. This like-suppresses-like motif reduced population redundancy, and modeling suggested a role assisting inference of the external features responsible for sensory inputs. Influence mapping might be extended to other areas, in order to verify or disprove hypotheses generated by the quantitative comparisons of neural representations employed in the first approach.
Citable link to this pagehttp://nrs.harvard.edu/urn-3:HUL.InstRepos:42029667
- FAS Theses and Dissertations