Neural Encoding and Decoding with Deep Learning for Natural Vision

Haiguang Wen, Purdue University

Abstract

The overarching objective of this work is to bridge neuroscience and artificial intelligence to ultimately build machines that learn, act, and think like humans. In the context of vision, the brain enables humans to readily make sense of the visual world, e.g. recognizing visual objects. Developing human-like machines requires understanding the working principles underlying the human vision. In this dissertation, I ask how the brain encodes and represents dynamic visual information from the outside world, whether brain activity can be directly decoded to reconstruct and categorize what a person is seeing, and whether neuroscience theory can be applied to artificial models to advance computer vision. To address these questions, I used deep neural networks (DNN) to establish encoding and decoding models for describing the relationships between the brain and the visual stimuli. Using the DNN, the encoding models were able to predict the functional magnetic resonance imaging (fMRI) responses throughout the visual cortex given video stimuli; the decoding models were able to reconstruct and categorize the visual stimuli based on fMRI activity. To further advance the DNN model, I have implemented a new bidirectional and recurrent neural network based on the predictive coding theory. As a theory in neuroscience, predictive coding explains the interaction among feedforward, feedback, and recurrent connections. The results showed that this brain-inspired model significantly outperforms feedforward-only DNNs in object recognition. These studies have positive impact on understanding the neural computations under human vision and improving computer vision with the knowledge from neuroscience.

Degree

Ph.D.

Advisors

Liu, Purdue University.

Subject Area

Biomedical engineering|Electrical engineering|Computer science

COinS