Project repository for the 2021 version of the Neuromorphic Computing course.
We tried to reproduce the spiking CNN of Tavanaei and Maida [2017] and extend it by changing the model architecture and hyperparameters.
The accuracies reported by the 2017 paper could not be reproduced, but we did find some interesting but unintuitive results when combining promising extensions. The number of active neurons of the STDP layer seems to have a large effect on the effectiveness of the model, but it is very difficult to control.
This repository uses the scikit-learn interfaces for training the models. The baseline model can be found in main.py, where most figures as shown in the paper are shown during training. A different architecture can easily be trained by adding or exchanging layers.
The models reported in the paper can be retrained with retrain.py and evaluated on the validation split with eval.py. Use pip install -r requirements.txt to install the requirements before running the code.
| Parameters | Accuracy |
|---|---|
| baseline | 0.9235 |
| filter_size=3 | 0.9170 |
| filter_size=7 | 0.9188 |
| filter_count=16 | 0.9215 |
| filter_count=64 | 0.9304 |
| memory=0.25 | 0.9459 |
| memory=0.368 | 0.9521 |
| memory=0.5 | 0.8952 |
| memory=0.75 | 0.9606 |
| memory=0.905 | 0.9570 |
| memory=1.0 | 0.9562 |
| norm=false | 0.4445 |
| lr=oja | 0.9407 |
| lr=bcm | 0.8546 |
| lr=stdp | 0.8767 |
| stdp_count=64 | 0.9015 |
| stdp_count=256 | 0.9235 |
| conv_conv-32,3 | 0.8759 |
| conv_conv-32,5 | 0.8004 |
| conv_conv-32,7 | 0.6343 |
| conv_init=norm | 0.9253 |
| conv_init=gluni | 0.9115 |
| conv_init=glnorm | 0.9207 |
| no conv/pool | 0.9308 |
| no stdp | 0.9824 |
| no conv/pool/stdp | 0.969 |