Figure 1: HaH block for image classification DNNs.
Official repository for the paper entitled "Towards Robust, Interpretable Neural Networks via Hebbian/anti-Hebbian Learning: A Software Framework for Training with Feature-Based Costs". If you have questions you can contact metehancekic [at] ucsb [dot] edu.
Maintainers: WCSL Lab, Metehan Cekic, Can Bakiskan,
numpy==1.20.2
torch==1.10.2
The most recent stable version can be installed via python package installer "pip", or you can clone it from the git page.
pip install hahtorch
or
git clone git@github.com:metehancekic/HaH.git
We used CIFAR-10 image classification to show the effectiveness of our module. We train a VGG16 in a standard fashion and train another VGG16 that contains HaHblocks with layer-wise HaHCost as a supplement. Details of our experiments can be found in our recent paper
Figure 2: HaH VGG16, our proposed architecture for HaH training, see paper for more detail.
Table 1: CIFAR10 classification: Performance of the HaH trained network against different input corruptions on the test set. For all of the adversarial attacks, we use AutoAttack which is an ensemble of parameter-free attacks, see paper for more detail.
0.0.5