Model Interpretability with Captum
summary
Library called Captum provides PyTorch gradient‑based attributions including integrated gradients, DeepLIFT, SmoothGrad and VarGrad and computes attributions by forward and backward passes (integral of gradients along a path and chain‑rule conductance for layers) with a few lines of code. It ships a visualization app Captum Insights that shows pixel and token attributions for ResNet, IMDB text, Titanic MLP and VQA examples. First release is gradient‑only, open source on github and captum.ai, with plans for more algorithms and benchmarks.
AI-written summary · about
chapters (4)
- 0:00Start
- 0:07Model Interpretability
- 1:50Gradient Based Attribution Algorithms
- 6:18Text Classification Model
related talks
27mUnderstanding Deep Neural Networks
35mDeep Learning Model Interpretability for Computer Vision based Models
32mPyTorch Demystified, Why Did I Switch
09mLightning Talk: Bridging the Gap: Engineering Compliant...
01h 04mFrom Pixels to Words: Exploring Multi-Modal AI with Python
45mScale Learning and Reasoning Across Heterogeneous Gradients and Semantics
46mIncreasing Participant Agency by Understanding the Impact of Training Data
48mInherent Trade-offs with the Local Explanations Paradigm