Network Introspection

Network introspectors help analyze model performance, offering suggestions for how to improve model efficiency and size.

Principal Filter Analysis

Guide neural network compression by removing redundant filters detected by correlated filter responses, to target a minimum loss in accuracy, or a specific model footprint.

Inactive Unit Analysis

Evaluates the extent to which units are inactive for a given set of probe data points by computing aggregate statistics using the unit responses to the input probes.