Explainable

DeepDebugger faithfully visualizes the training dynamics of any deep models and provide more insight to the execution trace of deep neural networks, thus making the fault localization of DNNs more explainable.

Fast and Efficient

DeepDebugger offers automatic and live-update visualization. This is essential to identify potential data anomalies, enabling timely interventions and conserving significant computational resources.

Tool Overview

DeepDebuuger records the training process of deep classifiers and projects the high-dimensional classification landscape into a two-dimensional space. Users can view the training dynamics of the high-dimensional classification landscape in the low-dimensional space, as a visualized animation. DeepDebuuger contributes to recommending user-interested samples in a human-in-the-loop manner. Given a debugging task requiring sample inspection, DeepDebuuger recommends what samples need to be (re)labelled or what samples are suspicious. Users can provide feedback on the recommendation, and DeepDebuuger can interactively recommend new samples, guiding the users to pinpoint the root cause of model bugs like misprediction.

Applications

Data Platform Visualization and Management

DeepDebugger provides a centralized environment to manage, process, and visualize data. This is a well-designed data platform combines data ingestion, storage, processing, and visualization tools to facilitate data-driven decision-making.

Data Flywheel

Deepdebugger designs a pipeline for machine learning models that can improve over time, creating a feedback loop that continuously enhances performance, data quality, and user experience.

DeepVisualInsight: A interpretable Visualizer for Deep Models.

DeepVisualInsight and TimeVis implements a novel visualizer for deep models.

Download.

Documentation.

Conference Paper.