An Empirical Analysis of Backward Compatibility in Machine Learning Systems

By Megha Srivastava, Besmira Nushi, Ece Kamar, Shital Shah, Eric Horvitz. 2020

53 citations recorded by the source corpus.

In many applications of machine learning (ML), updates are performed with the goal of enhancing model performance. However, current practices for updating models rely solely on isolated, aggregate performance analyses, overlooking important dependencies, expectations, and needs in real-world deployments. We consider how updates, intended to improve ML models, can introduce new errors that can significantly affect downstream systems and users. For example, updates in models used in cloud-based classification services, such as image recognition, can cause unexpected erroneous behavior in systems that make calls to the services. Prior work has shown the importance of "backward compatibility" for maintaining human trust. We study challenges with backward compatibility across different ML architectures and datasets, focusing on common settings including data shifts with structured noise and ML employed in inferential pipelines. Our results show that (i) compatibility issues arise even without data shift due to optimization stochasticity, (ii) training on large-scale noisy datasets often results in significant decreases in backward compatibility even when model accuracy increases, and (iii) distributions of incompatible points align with noise bias, motivating the need for compatibility aware de-noising and robustness methods.

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

Citations in the corpus

Forward Compatible Few-Shot Class-Incremental Learning (2022, background): “They are introduced to the machine learning field in [5, 38].”
Challenges in Deploying Machine Learning: A Survey of Case Studies (2020, background): “[132], who provide a detailed empirical study of backward compatibility issues in ML systems.”
Deep Class-Incremental Learning: A Survey (2023, unknown): “CIL with Bidirectional Compatibility: Compatibility is a design characteristic in software engineering community [240], [241], [242], [243], which was introduced to the machine learning community in [244], [245].”
Boundary-aware Backward-Compatible Representation via Adversarial Learning in Image Retrieval (2023, unknown): “However, the upgrade of CMC models still needs two embedding models and two databases which is shown in supplemental materials.”
Capabilities of GPT-4 on Medical Challenge Problems (2023, background): “The leukocyte count is still slightly elevated but closer to the normal range, and the eosinophils percentage has decreased, making an infectious or parasitic cause, such as schistosomiasis, less likely.”
Class-Incremental Learning: A Survey (2023, unknown): “CIL with Bidirectional Compatibility : Compatibility is a design characteristic in software engineering community [229], [230], [231], [232], which was introduced to the machine learning community in [233], [234].”
Are We Learning Yet? A Meta Review of Evaluation Failures Across Machine Learning (2021, background): “…to the benchmarking paradigm, such as test suites [1, 126, 170], for example, can re-orient ML evaluation away from its current focus on the competitive determination of “state of the art”, and more towards an exploratory and descriptive probing of model capabilities [65, 106, 143, 161, 169].”