OPEN SOURCE

ML APPLICATIONS

Captum and Fiddler partner to improve Explainable AI

July 13, 2020

AI has some explaining to do. As machine learning (ML) algorithms have found use cases in high-stakes areas ranging from health care to recruiting to banking and lending, the need to understand the decisions these algorithms are making is more important than ever. As part of a continuing effort to expand the tools available for Explainable AI, we’ve made Captum interoperable with Fiddler, the Explainable AI platform developed by Fiddler Labs.

Today, engineers can see the results for deployed ML algorithms, but their increasing sophistication makes understanding the how and why of their decisions very difficult. This creates issues of trust not only with developers but also with end users. Captum was introduced as part of PyTorch 1.3 as a tool to address this challenge. Captum utilizes state-of-the-art interpretability algorithms to examine ML models, allowing engineers to create more understandable, and thus more reliable and predictable, AI systems.

Fiddler lets engineers and developers analyze the decision-making behavior behind AI models and gather actionable insights. These insights enable them to safeguard against and resolve issues of mispredictions, bias, and unfairness. By combining Fiddler with Captum, PyTorch’s open source model interpretability library, we create a powerful toolset for research and applications into AI explainability.

By uploading explanations generated by Captum into Fiddler, developers can augment their ability to understand and predict the behavior of ML models. Once the Fiddler API has the model and its associated data set, it can analyze and debug the model and serve explanations at scale. In the case of a production model, Fiddler can also analyze data drift and outliers, as well as monitor live traffic for the model, among other services.

We believe that the interoperability between Captum and Fiddler will advance research and applications for AI explainability and encourage more engineers and developers to adopt Explainable AI in their ML workflows.

Learn more about the Fiddler and Captum collaboration to enhance Explainable AI offerings.