Center for the Neural Basis of Cognition

CNBC Postdoc Writing Group

Fridays, 2-4pm

Contact: Andrew Gerlach (arg151@pitt.edu)

Location is typically Zoom: https://pitt.zoom.us/j/5307316889 (Passcode: 1234)

Description: Two hour block dedicated to writing papers, grants, reviews, etc. We use the Pomodoro system of 25 min blocks with 5 min breaks in between to chitchat.

There’s a group of ~10 people who attend semi-regularly. On any given week, it’s typical to have 3-5 people. It’s used for accountability and setting aside a dedicated chunk of time for writing (although some people use it for analysis or whatever else they may need to focus on). It’s also been helpful in building a postdoc community.

Please reach out to Andrew with any questions!

Dear community,

Please see below for information regarding my thesis defense in two weeks. 

I hope to see many of you there in person, and a zoom option is available as well!

Presenter: Nicholas M. Blauch
Program: Program in Neural Computation, Carnegie Mellon Neuroscience Institute
Date: Monday, December 4
Time: 11 am
Location: Cohon University Center Danforth Conference Room, and Zoom: https://cmu.zoom.us/j/91595922356?pwd=QndTR0FxdXpnVzBXNm9QZThOUHNZQT09
Title: A connectivity-constrained computational account of topographic organization in high-level visual cortex
Committee: David C. Plaut (co-chair), Marlene Behrmann (co-chair), Leila Wehbe, Michael Arcaro (U Penn)
 
Abstract:
Visual recognition is an essential biological skill, allowing organisms to know where they are and what objects and individuals are available for interaction. In humans, a critical brain area for visual recognition is the ventral temporal cortex (VTC), a large extrastriate visual area situated towards the end of the ventral visual hierarchy, containing multiple cytoarchitectonic subdivisions. 
Neuroimaging and direct neural recordings have revealed substantial functional organization of VTC, including category-selective clusters (for faces, places, bodies, words, etc.) with a relatively consistent spatial layout across individuals. Damage to these regions can have highly specific effects, such as the loss of the ability to recognize faces or words, with much weaker effects on the recognition of other categories. These findings have been taken as evidence of an innately specified domain-specific organization, in which different categories are processed by distinctive high-level visual mechanisms. However, other viewpoints have emphasized the global organizational principles of VTC, such as the smooth mapping of object space dimensions, and other properties such as visual field eccentricity, “animacy” and the real-world size of objects. This has suggested that VTC may be better considered as a unitary domain-general mechanism with underlying specializations. 
This thesis aims to develop a computational account of functional organization in VTC from basic principles, and in so doing, explain the emergence of both domain-specific and domain-general aspects of organization, as well as aspects of their consistent global layout. The key idea is that such organization can emerge in neural networks optimized to perform tasks with inherent structure, under explicit constraints to minimize wiring costs. 
First, the ability of modern Deep Convolutional Neural Networks (DCNNs) to account for human expert face recognition behaviors is established, revealing the role of expert specialized features in face recognition, despite some overlap with more general features useful for recognizing other categories.
Second, a class of wiring-constrained recurrent neural networks (RNNs), termed interactive topographic networks (ITNs), is introduced, in which a hierarchy of RNN layers receives inputs from a DCNN that extracts performant visual features for multiple tasks. Several simulations are performed to demonstrate and understand the role of wiring and sign-based connectivity constraints on the formation of topographic organization in these models. Learned specialization is studied extensively, and insights are derived in relation to human and non-human primate cortical areas. 
Third, the ability of ITNs to develop topographic organization without semantic supervisory signals is demonstrated, highlighting the ability of structured visual inputs to drive substantial functional organization in a generic manner. 
Fourth, connectivity constraints within interactive topographic network models are generalized, allowing for input and output systems to shape the organization of the high-level visual system. Such constraints anchor the topography to particular locations, allowing for the simulation of consistent global organization, including hemispheric lateralization.
Last, an empirical approach is taken, asking how the lateralized cortical representations of words and faces interact across human subjects, demonstrating a surprising degree of both overlap in representation, as well as independence in laterality across domains, challenging previous suggestions that these domains compete for cortical real estate during development. Brain-wide analyses demonstrate coupled patterns of lateralization across cortical networks, and implicate the roles of long range connectivity in shaping individual differences in lateralization. 
While the connectivity-constrained computational account developed in this thesis remains a work in progress, it takes several steps towards unifying both domain-specific and domain-general views of VTC, and local and long-range connectivity constraints on its functional organization. Further development of models within this framework should provide further fine-grained insights into the development and nature of the organization of function in biological brains. 

best,
Nick