Towards Accountable AI: Hybrid Human-Machine Analyses for Characterizing System Failure

TL;DR

This paper introduces Pandora, a hybrid human-machine approach for detailed analysis and explanation of system failures in complex AI systems, improving understanding beyond traditional aggregate metrics.

Contribution

Pandora provides a novel set of tools combining human insights and system data to characterize failures in multi-component machine learning systems.

Findings

Pandora enables detailed failure analysis in image captioning systems.

Case study shows Pandora improves debugging and system understanding.

Hybrid methods reveal failure conditions related to input and architecture.

Abstract

As machine learning systems move from computer-science laboratories into the open world, their accountability becomes a high priority problem. Accountability requires deep understanding of system behavior and its failures. Current evaluation methods such as single-score error metrics and confusion matrices provide aggregate views of system performance that hide important shortcomings. Understanding details about failures is important for identifying pathways for refinement, communicating the reliability of systems in different settings, and for specifying appropriate human oversight and engagement. Characterization of failures and shortcomings is particularly complex for systems composed of multiple machine learned components. For such systems, existing evaluation methods have limited expressiveness in describing and explaining the relationship among input content, the internal states of system components, and final output quality. We present Pandora, a set of hybrid human-machine methods and tools for describing and explaining system failures. Pandora leverages both human and system-generated observations to summarize conditions of system malfunction with respect to the input content and system architecture. We share results of a case study with a machine learning pipeline for image captioning that show how detailed performance views can be beneficial for analysis and debugging.