Back
Understanding, Formally Characterizing, and Robustly Handling Real-World Distribution Shift (CMU PhD Thesis, 2024)
webkilthub.cmu.edu·kilthub.cmu.edu/articles/thesis/Understanding_Formally_Ch...
A 2024 CMU ML PhD thesis providing theoretical grounding for distribution shift robustness; relevant to AI safety researchers interested in formal guarantees, out-of-distribution generalization, and reliable deployment of ML systems.
Metadata
Importance: 55/100book chapterprimary source
Summary
Elan Rosenfeld's CMU PhD thesis develops theoretical and empirical foundations for handling distribution shift in ML systems, covering adversarial robustness certification, latent variable models of distribution shift using causal structure, and empirical analysis of real-world data variation. It introduces the concept of 'environment/intervention complexity' as a core measure for domain generalization and causal representation learning.
Key Points
- •Proposes scalable methods for certifying deep neural network robustness to adversarial attacks on test samples, training data, or any model-influencing input.
- •Develops latent variable models of distribution shift grounded in causality, enabling formal analysis of multi-distribution robust learning methods.
- •Introduces 'environment/intervention complexity' as a statistical measure quantifying identifiability conditions for domain generalization.
- •Empirically investigates real-world heavy tails and distribution shift to understand practical failures of modern ML systems.
- •Argues benchmarks fundamentally cannot capture all real-world variation, motivating formal characterizations of shift structure.
Cited by 1 page
| Page | Type | Quality |
|---|---|---|
| AI Distributional Shift | Risk | 91.0 |
Cached Content Preview
HTTP 200Fetched Apr 7, 20265 KB
Item - Understanding, Formally Characterizing, and Robustly Handling Real-World Distribution Shift - Carnegie Mellon University - Figshare Understanding, Formally Characterizing, and Robustly Handling Real-World Distribution Shift
Expand https://doi.org/10.1184/R1/26312050 Copy identifier URL to clipboard Identifier Info
Cite Download ( 12.74 MB ) Share Embed thesis posted on 2024-07-23, 16:42 authored by Elan Rosenfeld Elan Rosenfeld <p>Distribution shift remains a significant obstacle to successful and  reliable deployment of machine learning (ML) systems. Long-term  solutions to these vulnerabilities can only come with the understanding  that benchmarks fundamentally cannot capture all possible variation  which may occur; equally important, however, is careful experimenta tion with AI systems to understand their failures under shift in practice.  </p> <p>This thesis describes my work towards building a foundation for trustworthy and reliable machine learning. The surveyed work falls  roughly into three major categories: (i) designing formal, practical char acterizations of the structure of real-world distribution shift; (ii) leverag ing this structure to develop provably correct and efficient learning algo rithms which handle such shifts robustly; and (iii) experimenting with  modern ML systems to to understand the practical implications of real world heavy tails and distribution shift, both average- and worst-case.  </p> <p>Part I describes work on scalably certifying the robustness of deep  neural networks to adversarial attacks. The proposed approach can  be used to certify robustness to attacks on test samples, training data,  or more generally any input which influences the model’s eventual  prediction. In Part II, we focus on latent variable models of shifts,  drawing on concepts from causality and other structured encodings of  real-world variation. We demonstrate how these models enable for mal&#
... (truncated, 5 KB total)Resource ID:
56f1ba822bd9862d | Stable ID: sid_n59mef9SzE