tslens: A PyTorch Framework for Interpreting Time Series Deep Learning Models

Résumé fourni par la source

tslens: A PyTorch Framework for Interpreting Time Series Deep Learning Models tslens is a drop-in, Captum-compatible interpretability toolkit for PyTorch time-series models, spanning supervised and foundation models. Every method returns a saliency map over the input window, pinpointing when and where a model found evidence for its prediction. Interpretation methods built for vision and NLP don't transfer cleanly to time series: neighbouring time steps are strongly correlated, and a feature's importance can shift from one time step to the next. tslens gathers over a dozen saliency and attribution methods designed for or adapted to this setting, some native, others wired in from Captum and Time Interpret.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.