IEEE VIS 2024 Content: DETOXER: A Visual Debugging Tool With Multiscope Explanations for Temporal Multilabel Classification

DETOXER: A Visual Debugging Tool With Multiscope Explanations for Temporal Multilabel Classification

Mahsan Nourani -

Chiradeep Roy -

Donald R. Honeycutt -

Eric D. Ragan -

Vibhav Gogate -

Access paper PDF
DOI: 10.1109/MCG.2022.3201465

Room: Bayshore III

2024-10-16T16:24:00ZGMT-0600Change your timezone on the schedule page
2024-10-16T16:24:00Z
Exemplar figure, described by caption below
Overview of DETOXER, a visual (de)bugging (to)ol with Multi-Scope E(x)planations for (er)ror detection in Temporal Multi-Label Classification. In the center, a video is selected for exploration. Directly under the progress bar, heatmaps demonstrate the model’s confidence for any given label per second (frame-level explanations)-(C). On the left, available videos are shown; for each video, the tool shows top-5 detected labels (A) and the rate of FP and FN errors (B) in the video (video-level explanations). The selected video is emphasized with a blue background. On the right, a global information panel displays model performance metrics (D) and object-specific FN and FP error rates in two vertically adjacent bar charts (E) (Global-level explanations).
Fast forward
Full Video
Keywords

Debugging, Analytical Models, Heating Systems, Data Models, Computational Modeling, Activity Recognition, Deep Learning, Multi Label Classification, Visualization Tool, Temporal Classification, Visual Debugging, False Positive, False Negative, Active Components, Deep Learning Models, Types Of Errors, Video Frames, Error Detection, Detection Of Types, Action Recognition, Interactive Visualization, Sequence Of Points, Design Goals, Positive Errors, Critical Outcomes, Error Patterns, Global Panel, False Negative Rate, False Positive Rate, Heatmap, Visual Approach, Truth Labels, True Positive, Confidence Score, Anomaly Detection, Interface Elements

Abstract

In many applications, developed deep-learning models need to be iteratively debugged and refined to improve the model efficiency over time. Debugging some models, such as temporal multilabel classification (TMLC) where each data point can simultaneously belong to multiple classes, can be especially more challenging due to the complexity of the analysis and instances that need to be reviewed. In this article, focusing on video activity recognition as an application of TMLC, we propose DETOXER, an interactive visual debugging system to support finding different error types and scopes through providing multiscope explanations.