Boston University
Improving performance, efficiency, and stability in self-supervised machine learning
Abstract
dc:description.abstractThis dissertation advances self-supervised learning procedures as a principled response to the labeled data bottleneck, a fundamental and pervasive constraint in applied machine learning. In many real-world settings, acquiring sufficient labeled examples is prohibitively expensive or logistically infeasible, yet standard supervised approaches offer little recourse. By leveraging multi-agent reinforcement learning and adversarial training, models can learn from richer, less curated signals, reducing dependence on hand-labeled data while preserving task-relevant inductive biases. First, we introduce a biologically plausible complex-valued neural network architecture inspired by neural synchrony in the brain. A novel activation function enables neurons to respond preferentially to temporally synchronized inputs, yielding improved performance on benchmark datasets compared to real-valued fully connected networks. We demonstrate how complex-valued neural architecture can improve the performance of adversarial learning algorithms without breaking the delicate balance between adversarial networks. Second, we develop a learning framework for computing coarse correlated equilibria (CCE) in large, simultaneous-move games. The method combines no-regret learning, self-play, and deep neural networks to approximate CCE values and iteratively guide a tree search process. The resulting algorithm generalizes well across game states and outperforms existing multi-agent reinforcement learning baselines. Third, we propose a modified Wasserstein GAN (WGAN) for reweighting biased survey data to match a reference population. Instead of generating synthetic data, the generator learns sample weights that minimize distributional discrepancy between the biased and reference datasets. Applied to real-world surveys, the model recovers richer variable interactions and achieves superior predictive performance on outcomes such as COVID-19 vaccination and insurance rates. Building on the survey reweighting framework, we further improve performance by incorporating complex-valued network components and applying the no-regret stopping criterion from the second project, with both advances yielding measurably better reweighting. Motivated by the success of no-regret stopping in adversarial training, we then apply this criterion to an entirely different domain: unsupervised domain adaptation via Domain Adversarial Neural Networks (DANN). Despite the architectural differences between WGANs and DANN, the no-regret stopping criterion transfers naturally. Both WGANs and DANN share the fundamental structure of a minimax adversarial game, making the transfer principled. Applied to DANN, the no-regret stopping criterion alone produces statistically significant improvements in target domain generalization across multiple benchmarks, confirmed by statistical tests of significance. Together, these contributions demonstrate that aligning neural learning systems with the underlying structure of their domain leads to more robust, efficient, and accurate models. This thesis advocates for a domain-informed approach to neural network design, enabling better solutions to complex applied problems across biology, game theory, and statistical inference.
Degree
thesis:*- Name thesis:degree_name
- Doctor of Philosophy
- Level thesis:degree_level
- doctoral
- Discipline thesis:degree_discipline
- Computer Science
- Grantor
- Boston University
- Year dc:date.issued
- 2026
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Yu, Ryan
- Advisor dc:contributor.advisor
-
- Olshevsky, Alexander
Subjects
dc:subject × 4Rights
dc:rights- Statement dc:rights
-
- Attribution 4.0 International
- Licence dc:rights.uri
- Language dc:language.iso
- en_US
Identifiers
dc:identifier.*- Handle dc:identifier.uri
- https://hdl.handle.net/2144/53440