2023 Theses Doctoral

Complaint Driven Training Data Debugging for Machine Learning Workflows

Flokas, Lampros

As the need for machine learning (ML) increases rapidly across all industry sectors, so has theinterest in building ML platforms that manage and automate parts of the ML life-cycle. This has enabled companies to use ML inference as a part of their downstream analytics or their applications. Unfortunately, debugging unexpected outcomes in the result of these ML workflows remains a necessary but difficult task of the ML life-cycle. The challenge of debugging ML workflows is that it requires reasoning about the correctness of the workflow logic, the datasets used for inference and training, the models, and interactions between them. Even if the workflow logic is correct, errors in the data used across the ML workflow can still lead to wrong outcomes. In short, developers are not just debugging the code, but also the data.

We advocate in favor of a complaint driven approach towards specifying and debugging data errors in ML workflows. The approach takes as input user specified complaints specified as constraints over the final or intermediate outputs of workflows that use trained ML models. The approach outputs explanations in the form of specific operator(s) or data subsets, and how they may be changed to address the constraint violations.

In this thesis we make the first steps towards our complaint driven approach to data debugging. As a stepping stone, we focus our attention on complaints specified on top of relational workflows that use ML model inference and whose errors are caused by errors in ML model’s training data. To the best of our knowledge, we contribute the first debugging system for this task, which we call Rain. In response to a user complaint, Rain ranks the ML model’s training examples based on their ability to address the user’s complaint if they were removed. Our experiments show that users can use Rain to debug training data errors by specifying complaints over aggregations of model predictions without having to specify the correct label for each individual prediction.

Unfortunately, Rain’s latency may be prohibitive for use in interactive applications like analytical dashboards or business intelligence tools where users are likely to observe errors and complain. To address Rain’s latency problem when scaling to large ML models and training sets, we propose Rain++. Rain++ pushes the majority of Rain’s computation offline ahead of user interaction, achieving orders of magnitude online latency improvements compared to Rain.

To go beyond Rain’s and Rain++’s approach that evaluates individual training example deletionsindependently we propose MetaRain, a framework for training classifiers that detect training data corruptions in response to user complaints. Thanks to the generality of MetaRain, users can adapt the classifiers chosen to the training corruptions and the complaints they seek to resolve. Our experiments indicate that making use of this ability results in improved debugging outcomes.

Last but not least, we study the problem of updating relational workflow results in response tochanges to the inference ML model used. This can be leveraged by current or future complaint driven debugging systems that repeatedly change the model and reevaluate the relational workflow. We propose FaDE, a compiler that generates efficient code for the workflow update problem by casting it as view maintenance under input tuple deletions. Our experiments indicate that the code generated by FaDE has orders of magnitude lower latency than existing view maintenance systems.