Title

Feature Enhancement with Deep Generative Models In Deep Bayesian Active Learning

Abstract

Abstract

Data-intensive models emerge as new advances in Deep Learning take place. However, access to annotated datasets with many data points are not constantly prevalent. This situation emphasizes the need for Active Learning to select the least possible amount of data without compromising the accuracy of the models. Recent advancements occur in Deep Bayesian Active Learning (DBAL), which means incorporating uncertainty of model parameters into a Deep Network. In this work, we present an algorithm that improves the accuracy of a DBAL model in an image classification task. We utilize the representation power of Deep Generative Models by employing their feature extraction capabilities. We obtain improved feature space representation of input data referred to as a latent vector by training a generative model. Instead of using the entire image space in the active learning setting, we demonstrate that utilizing latent space provides better data point selection for the active learning problem, hence obtaining higher accuracy. We validate that the proposed method outperforms the baseline model on the benchmark datasets. Furthermore, this study compares different generative models in terms of the ability to capture the better feature representation. Informativeness of the data points defines how well an active learning algorithm performs. Therefore, capturing the latent space representation of a data point by extracting the highest information value possible is a significant contribution. We provide comparisons and experiments on different kinds of Generative Models, namely Vanilla Variational Autoencoders (VAEs), Maximum Mean Discrepancy Variatioanal Autoencoders (MMDVAE) and Bidirectional Generative Adversarial Networks (BiGANs). Additionally, Bayesian Active Learning suffers from Mode-Collapse problem. In order to ease that, we propose a diversity-based query algorithm and enhance the diversity of the active points.

Supervisor(s)

Supervisor(s)

PINAR EZGI DUYMUS

Date and Location

Date and Location

2022-09-01 14:00:00

Category

Category

MSc_Thesis