Deep Feature Factorization
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Type Name Latest commit message Commit time
Failed to load latest commit information.


Deep Feature Factorization (DFF)

DFF is the application of non-negative matrix faxtorization (NMF) to the ReLU feature activations of a deep neural network. In the case of CNNs trained on images, the resulting factors decompose an image batch into semenatic parts with a high degree of invariance to complex transformations.

This implementation relies on Pytorch and includes a GPU implementation of NMF with multiplicative updates (Lee and Seung, 2001).

author = {Collins, Edo and Achanta, Radhakrishna and Susstrunk, Sabine}, title = {Deep Feature Factorization For Concept Discovery}, booktitle = {The European Conference on Computer Vision (ECCV)}, year = {2018}