Examination Committee

Prof Dekai WU, CSE/HKUST (Chairperson)
Prof Bertram SHI, ECE/HKUST (Thesis Supervisor)
Prof André VAN SCHAIK, The MARCS Institute for Brain, Behaviour and Development, Western Sydney University (External Examiner)
Prof Pascale FUNG, ECE/HKUST
Prof Matthew MCKAY, ECE/HKUST
Prof Richard H Y SO, IELM/HKUST

Abstract

The human visual system is still leagues ahead of the computer vision algorithms introduced within the last few decades. One main reason behind this is that the brain is able to adapt based on the statistics of the sensory inputs. Understanding the computational principles behind this learning process is important in designing adaptive artificial agents. 

Sparsity and temporal slowness have been identified as two critical components in shaping visual receptive fields in the primary visual cortex. Sparsity constraint posits that neural population responses represent sensory data using as few active neurons as possible. Temporal slowness assumes that neurons adapt to encode information about the environment, which is relatively stable in comparison to the raw sensory signals. Although both slowness and sparsity might lead to the development of invariant feature detectors, there is no clear agreement on their relative contributions in neuronal development.  In this thesis, we propose the Generative Adaptive Subspace Self Organizing Map (GASSOM), which utilizes sparsity and slowness in learning invariant feature extractors. Using this framework, we show that temporal slowness can emerge in the model as it tries to learn a better representation of sensory signals, and that incorporating slowness result in representations that exhibit better invariance. We validate the applicability of the GASSOM framework in tasks that require the learning of invariant visual representations. 

We show that the GASSOM can be used in developmental robotics, and use it for explaining neurophysiological findings in rodent vision. We show the applicability of the GASSOM as a generic learning algorithm for hierarchical organizations of feature extractors that model the information flow in the visual cortex. Finally we extend the GASSOM to the event domain, by constructing a framework for learning invariant feature detectors from stimuli generated using event-driven neuromorphic vision sensors.