Title: Deep Dreaming for Few-Shot Learning in Neural Networks
Presenter: Eugene D. Hanson
Advisor: Dr. Vangelis Metsis
Date/Time: May 10, 11:00 a.m.
Location: https://txstate.zoom.us/j/93926488938
Abstract:
Few-shot learning is a machine learning method by which a trained model learns to recognize a new class of data by looking at only a few in-stances of that class. Existing methods for few-shot learning rely on data augmentation and contrastive learning algorithms to fine-tune pre-trained models for the new class added. However, when creating augmented versions of the few new shots, existing methods do not take into consideration what the pre-trained model already knows. Our method uses a deep dream-like generator to augment the instances of the new class with information that the network has learned in the past. The generator over-interprets and enhances the known patterns it sees in an instance of the new class. When the augmented instance is then fed back to the model for training, the model learns to ignore those patterns and focus on new features that are representative of the new class. This process results in more efficient and effective few-shot training for deep learning models.
Deadline: June 8, 2022, midnight