ICLM Journal Club
This Friday - January 20, 2023 (9:30 am, in person, Gonda 1357)
Speaker: Gergo Orban
Title: Top-down inference in a hierarchical account of the early visual cortex
Summary: Learning hierarchical representations from natural images without supervision is a central element of the theory of biological vision. The study of generative models and natural images have provided invaluable insights into the functioning of low level vision and in particular V1. Receptive field structure, response nonlinearities, noise correlations, and variability of population responses are all related to different aspects of probabilistic inference in generative models. Going beyond V1, however, has been hampered by a lack of understanding how learning and inference in nonlinear generative models can be performed. Recently, a flexible class of generative models, Variational Autoencoders (VAEs) have been proposed which can perform these by performing variational inference using a pair of jointly trained models, the recognition and generative models. However, training hierarchical VAEs has proven a daunting task. We recruit insights from neuroscience to shed light on how architectural choices and inductive biases contribute to shaping hierarchical representations and hierarchical computations. We introduce TD-VAE, a two hidden layer VAE that reflects the architectural properties of V1 and V2 of the visual cortex. We use natural image patches for training in order to explore the capabilities of TD-VAE. We identify that gradual compression along the cortical hierarchy leads to sensitivities at the level of V2 in the model reminiscent in many ways those found in electrophysiological recordings in nonhuman primates. Further, we identify a principled formulation of hierarchical inference which provides a normative framework to interpret perceptual top-down effects in mean responses and response correlations at the level of V1. In summary, VAEs are capable of accounting for a range of representational and coding properties in V1 and V2 and we propose hierarchical VAEs as alternatives to discriminative models to investigate coding properties of sensory cortical areas.
Relevant Papers:
About Us
Introduction
The Integrative Center for Learning and Memory (ICLM) is a multidisciplinary center of UCLA labs devoted to understanding the neural basis of learning and memory and its disorders. This will require a unified approach across different levels of analysis, including;
1. Elucidating the molecular cellular and systems mechanisms that allow neurons and synapses to undergo the long-term changes that ultimately correspond to 'neural memories'.
2. Understanding how functional dynamics and computations emerge from complex circuits of neurons, and how plasticity governs these processes.
3. Describing the neural systems in which different forms of learning and memory take place, and how these systems interact to ultimately generate behavior and cognition.
History of ICLM
The Integrative Center for Learning and Memory formally LMP started in its current form in 1998, and has served as a platform for many interactions and collaborations within UCLA. A key event organized by the group is the weekly ICLM Journal Club. For more than 10 years, graduate students, postdocs, principal investigators, and invited speakers have presented on topics ranging from the molecular mechanisms of synaptic plasticity, through computational models of learning, to behavior and cognition. Dean Buonomano oversees the ICLM journal club with help of student/post doctoral organizers. For other events organized by ICLM go to http://www.iclm.ucla.edu/Events.html.
Current Organizers:
Saray Soldado (Buonomano Lab) & Lukas Oesch (Churchland Lab). Please email us at iclm.journalclub@gmail.com if you would like to get regular updates regarding our journal club and weekly reminders.
Current Faculty Advisor:
Past Organizers:
i) Anna Matynia(Aug 2004 - Jun 2008) (Silva Lab)
ii) Robert Brown (Aug 2008 - Jun 2009) (Balleine Lab)
iii) Balaji Jayaprakash (Aug 2008 - Nov 2011) (Silva Lab)
iv) Justin Shobe & Thomas Rogerson (Dec 2011 - June 2013) (Silva Lab)
v) Walt Babiec (O'Dell Lab) (2013-2014)
vi) Walt Babiec (O'Dell Lab) & Helen Motanis (Buonomano Lab) (2014-2017)
vii) Helen Motanis (Buonomano Lab) & Shonali Dhingra (Mehta Lab) (2017-2018)
viii) Shonali Dhingra (Mehta Lab) (2018-2020)
ix) Megha Sehgal (Silva Lab) & Giselle Fernandes (Silva Lab) (2020-2022)
Wiki Newbies
Consult the User's Guide for information on using the wiki software.
