Deep Generative Classifier with Short Run Inference

Center for Vision, Cognition, Learning, and Autonomy at UCLA

I submitted this Masters thesis while advised by Dr. Song-Chun Zhu at the Center for Vision, Cognition, Learning, and Autonomy at UCLA.

A deep generative classifier employs Short Run Markov Chain Monte Carlo inference with Langevin dynamics and backpropagation through time. In contrast to a convolutional neural network (ConvNet) with analogous architecture, the Short Run classifier approaches the same classification accuracy and (1) may synthesize data, (2) may learn unsupervised from additional unannotated data, and (3) exhibits robustness to adversarial attacks, due to the stochasticity of the Langevin equation and the top-down architecture of the generator network. The ConvNet classifier lacks the ability to perform (1) or (2) and possesses no defense against adversarial attacks, a critical concern for any deployed machine learning system. Meanwhile, the Short Run classifier demonstrates the capacity to improve in both classification accuracy and the quality of data synthesis, given additional unannotated data.

Avatar
Eric Fischer
Statistics Ph.D. student, Generative Modeling, Representation Learning

I carry out research in generative modeling and representation learning for language.

Related