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).
- @InProceedings{collins2018,
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}
}