Towards The Deep Model : Understanding Visual Recognition Through Computational Models

Understanding how visual recognition is achieved in the human brain is one of the most fundamental questions in vision research. In this thesis I seek to tackle this problem from a neurocomputational modeling perspective. More specifically, I build machine learning-based models to simulate and explain cognitive phenomena related to human visual recognition, and I improve computational models using brain-inspired principles to excel at computer vision tasks.I first describe how a neurocomputational model (“The Model”, TM, (Cottrell & Hsiao, 2011)) can be applied to explain the modulation of visual experience on the performance of subordinate-level face and object recognition. Next, by introducing a mixture-of-experts structure in the model, I show that TM can be used to simulate the development of hemispheric lateralization of face processing. In addition, I extend TM to “The Deep Model” (TDM) by coupling it with deep learning techniques, and use TDM to explain the peripheral vision advantage in human scene recognition.Furthermore, I show the performance of these computational methods can be improved by introducing realistic constraints based on the human brain. By combining unsupervised feature learning principles with the Gnostic Fields theory of how the brain performs object recognition across the ventral visual pathway, I show a biologically-inspired model can develop realistic features of the early visual cortex, while performing well on object recognition datasets. By designing better encoding and decoding strategies in the deep neural network, I demonstrate that our system achieves the state-of-the-art performance on pixel-level semantic segmentation task on many popular computer vision benchmarks