Tuesday, May 02, 2023 - 02:00pm to Tuesday, May 02, 2023 - 04:00pm
NCS 220
Event Description

Machine learning has been highly successful in data-intensive applications, but is often hampered when the data set is small.Recently, few-shot learning (FSL) has been proposed to tackle this problem. It can be of great use in scenarios where data with supervised information are hard or impossible to acquire.

FSL aims to develop methods that can rapidly generalize to new tasks using very few samples with labels.In this thesis proposal, we explore the use of feature augmentation techniques on FSL problems. We aim to use generative models to model the feature distribution via a continuous space from which we can sample new data for augmentation and generation. We demonstrate thatwe can use this framework for various purposes, including generating representative features, transferring information between features, generating more difficult features and generating object templates. Our proposed solutions based on this framework achieve state-of-the-art performance for four different tasks: few-shot image classification, fine-grained few-shot classification, few-shot object detection and class-agnostic object counting. Specifically, for few-shot image classification, we propose a method for representative sample generation via sample selection. For fine-grained few-shot classification, we learn a distribution of intra-class variance to diversify features in a reliable manner. For few-shot object detection, we propose a novel feature generation system where we can control a specific property of the generated samples. For object counting, we use feature generation to introduce a zero-shot counter, a system that can count any object category, given the class name as the only input.

To conclude this thesis, we investigate an important aspect of generative models: how to automatically evaluate the quality of the generated samples.We observe that latent codes located in high-density regions of the latent manifold tend to produce high-quality generated samples, while codes in low-density areas often result in poor-quality ones. Building on this observation, we develop a metric to estimate the quality of generated outputs, in which a higher score indicates a lower likelihood of noise or artifacts. We propose to incorporate this quality score into our existing FSL frameworks using variational autoencoders and further, investigate the use of this metric in different generative frameworks.

Event Title
Ph.D. Thesis Defense: Jingyi Xu, 'Feature Generation and Augmentation for Few-Shot Learning'