This library implements the algorithms describe in the paper: Thoret, E., Andrillon, T., Leger, D., Pressnitzer, D. (2021) Probing machine-learning classifiers using noise, bubbles, and reverse correlation, Journal of Neuroscience Methods, 362, 109297 10.1016/j.jneumeth.2021.109297 10.1101/2020.06.22.165688 (pre-print)
It allows to probe the discriminative and canonical information of any trained machine learning classifier.
- Download or clone the repository.
- All the functions are in the
'./lib/proise.py'file and can be called by importinglibwith'from lib import proise'
A use case on the database of handwritten digits MNIST is provided in ./demo.py and summarized in below. The CNN trained on MNIST can be refitted with './fit_cnn_mnist.py'
The typical use of the reverse correlation method (Ahumada et al., 1976) uses the two following functions:
- Generate the probing samples:
proise.generateProbingSamples - Compute the canonical maps from the classifier predictions:
proise.ComputeCanonicalMaps
For MNIST, it provides the following output:
(see demo_mnist.py for a detailed implementation)
The typical use of the Bubbles method (Gosselin & Shyns, 2001) uses the three following functions:
- Generate bubble masks:
proise.generateBubbleMask - Generate the probing samples:
proise.generateProbingSamples - Compute the discriminative maps from the classifier predictions:
proise.ComputeDiscriminativeMaps
For MNIST, it provides the following output:
(see demo_mnist.py for a detailed implementation)
tensorflow, numpy, matplotlib, keras, tensorflow, sklearn, random
Ahumada Jr, A., & Lovell, J. (1971). Stimulus features in signal detection. The Journal of the Acoustical Society of America, 49(6B), 1751-1756.Gosselin, F., & Schyns, P. G. (2001). Bubbles: a technique to reveal the use of information in recognition tasks. Vision research, 41(17), 2261-2271.
- Developped by Etienne Thoret - For any questions/suggestions/bugs: please contact me at
firstnamename@gmail.com