Trustworthy Machine Learning under Distribution Shifts

TL;DR

This work addresses the reliability and trustworthiness of machine learning under real-world distribution shifts by proposing an integrated toolkit of methods. HOOD disentangles content and style to harness benign OOD data while detecting malign OOD data, improving OOD detection, SSL, and DA. SharpDRO introduces a sharpness-aware robust optimization to flatten loss landscapes under severe corruptions, with distribution-aware and distribution-agnostic variants. EVIL leverages sparse training to identify a robust invariant subnetwork by actively exploring less stable, variant parameters, boosting OOD generalization across tasks. Machine Vision Therapy (MVT) and Out-of-Modal Generalization (OOM) extend robustness to vision-language alignment and unseen modalities via denoising in-context learning and a Connect&Explore framework, with theoretical guarantees for convergence and optimality. Together, these chapters offer a comprehensive framework for robust, explainable, and adaptable AI across perturbations, domain shifts, and modality gaps, with broad implications for safety and deployment in high-stakes settings.

Abstract

Machine Learning (ML) has been a foundational topic in artificial intelligence (AI), providing both theoretical groundwork and practical tools for its exciting advancements. From ResNet for visual recognition to Transformer for vision-language alignment, the AI models have achieved superior capability to humans. Furthermore, the scaling law has enabled AI to initially develop general intelligence, as demonstrated by Large Language Models (LLMs). To this stage, AI has had an enormous influence on society and yet still keeps shaping the future for humanity. However, distribution shift remains a persistent ``Achilles' heel'', fundamentally limiting the reliability and general usefulness of ML systems. Moreover, generalization under distribution shift would also cause trust issues for AIs. Motivated by these challenges, my research focuses on \textit{Trustworthy Machine Learning under Distribution Shifts}, with the goal of expanding AI's robustness, versatility, as well as its responsibility and reliability. We carefully study the three common distribution shifts into: (1) Perturbation Shift, (2) Domain Shift, and (3) Modality Shift. For all scenarios, we also rigorously investigate trustworthiness via three aspects: (1) Robustness, (2) Explainability, and (3) Adaptability. Based on these dimensions, we propose effective solutions and fundamental insights, meanwhile aiming to enhance the critical ML problems, such as efficiency, adaptability, and safety.

Figures (36)

• Figure 1: Outline for trustworthy machine learning under distribution shifts. We explore Robustness, Explainability, and Adaptability for trustworthiness of ML models, and consider three types of common distribution shifts, including Perturbation Shift, Domain Shift, and Modality Shift. From Chapter 2 to Chapter 6, we shed light on different trustworthy aspects and distribution shift scenarios by solving realistic tasks and proposing effective approaches with theoretical grounding.
• Figure 2: (a) An ideal causal diagram denoting the data generating process. (b) Illustration of our disentanglement. The brown-edged variables $\tilde{C}$ and $\tilde{S}$ are approximations of content $C$ and style $S$. The dashed lines indicate the unwanted causal relations to be broken. (c) Illustration of the data augmentation of HOOD. The green lines and red lines denote the augmentation of benign OOD data $\bar{X}$ and malign OOD data $\hat{X}$, respectively. In all figures, the blank variables are observable and the shaded variables are latent.
• Figure 3: Architecture of the HOOD. The solid lines denote the inference flow, the dashed lines indicate the disentanglement of content and style, and the tildes stand for the approximation of the corresponding variables.
• Figure 4: Left: Augmentation number analysis. Right: CIFAR10 Visualization of our data augmentation.
• Figure 5: Illustration content ande style disentanglement on CIFAR10. The number in each cell denotes the prediction probability.

Theorems & Definitions (24)

• Theorem 1: Informal
• Proposition 2
• Theorem 4
• lemma 1
• Theorem 5
• Definition 6: Modality disagreement
• Theorem 7
• Remark 9
• Remark 11
• Definition 13: Stationary measure