Knowledge discovery and transfer on limited datasets

Abstract

Deep learning has achieved great success in many real-world applications, e.g., computer vision and human healthcare. Current deep learning usually relies on large-scale datasets with large amounts of carefully annotated data, however, objects in the real world follow a long-tailed distribution, i.e., a tremendous number of classes have little data and hard to be collected. Besides, new classes keep emerging that collecting, and particularly annotating examples is impossible. The data insufficiency poses a bottleneck in the robustness of deep learning methods.

Targeting this challenge, I propose to ease data insufficiency in two general data situations: single dataset and multiple datasets. For a single dataset, I propose to discover extra knowledge to enrich the learnable information for models (i.e., knowledge discovery). For multiple datasets, I propose to transfer generalizable knowledge to enhance the analysis of limited datasets (i.e., knowledge transfer).

To discover extra new knowledge in a single dataset, I propose to implement the discovery process in three levels. 1) Dataset level. Using contrastive relationships as new information to improve data efficiency on small-scale datasets. 2) Subject/class level. Enabling models to be aware about class/subject and thus be robust handling class/subject variance. 3) Instance level. Automating expert analysis as a new supervisor to boost model training.

To enable models to transfer knowledge between datasets, I propose to tackle three common transfer problems in the real world. 1) Lacking data in one of the datasets. I transfer cross-domain knowledge from sufficient to insufficient information domains to augment new data and ease data inefficiency. 2) Multi-view aggregation problem. I propose a general multi-view algorithm to transfer and unite different information from hierarchical views. 3) Biased knowledge transfer among datasets. I propose a task-aligned meta-learning model to learn generalizable knowledge for transferring.

In this thesis, I have conducted extensive experiments and ablation studies to validate the proposed methods on diverse real-world applications, e.g., limited-scale datasets of human activity recognition, image classification, neural signal analysis, and commercial analysis.