liveBook · Manning

Interpretability

As explained in Deep Learning with R, Third Edition

Interpretability is the degree to which a machine-learning model's learned representations, predictions, and decisions can be understood by humans.

Interpretability is the ability to understand how a machine-learning model represents inputs and reaches predictions. For image classifiers, it helps answer why a model identified an object, especially when supporting human expertise or investigating mistakes. section 10.0

Convolutional-network interpretability uses three complementary views:

  • Intermediate activations show how successive layers transform an image and suggest what individual filters detect. section 10.0
  • Filter visualization reveals the visual patterns or concepts to which particular filters respond. section 10.0
  • Class activation heatmaps identify image regions that contribute to a selected class prediction. section 10.0 section 10.3

Explaining an elephant prediction with Grad-CAM

A pretrained Xception model predicts an elephant photograph as African elephant with probability 0.903. Grad-CAM takes the final convolutional feature map, weights each channel by the class gradient, and produces a spatial heatmap showing which regions support the prediction. The same approach can help inspect a classification mistake or localize an object. section 10.3

How Deep Learning with R, Third Edition covers it

How deep it goes
Explains three visualization methods and the channel-weighting mechanics of Grad-CAM, while stopping short of surveying every technique.
Examples
Uses a dogs-versus-cats classifier, a pretrained Xception model, and an African-elephant image to demonstrate visual explanations.
What it assumes
Assumes access to a convolutional image classifier and input image while introducing the visualization techniques needed to inspect its decisions.
What you can do after
Diagnose/debug with it by visualizing layer activations, inspecting filters, and locating image regions behind a class prediction.
How it connects
Links convolutional neural networks with intermediate activations, filters, feature maps, class activation maps, Grad-CAM, and model debugging.
Cover of Deep Learning with R, Third Edition
liveBookmanning.com
Read Deep Learning with R, Third Edition on liveBook Get the book at manning.com

Also explained in

Compare how 16 books explain interpretability →

Related concepts

This explainer was generated by Manning's knowledge graph from the text of Deep Learning with R, Third Edition. The section links point to the parts of the book it draws on.