In the domain of AI & ML, images represent one of the most common and widely investigated type of data. Regardless of the specific field they belong to, images carry information hidden in their own spatial correlation.
Well-established models like CNNs can be successfully employed to disclose the relevant patterns that build such correlation.
The same idea can be applied to different data like gene expression, which might not come in form of image but that nevertheless carry information through an intrinsic spatial correlation.
Furthermore, as neuromorphic sensors have shown, this strategy can be even improved by focusing on sparse events rather than on a continuous exploitation of the whole amount of data.
In terms of software implementation, this translates into the adoption of neuro-inspired spiking CNNs, able to achieve state-of-the-art results with much smaller energy consumption.
- Clone the repository
- Install the requirements
- Move to
nni_configs
folder - Select the configuration file for the experiment you wish to perform, and then execute the following command:
nnictl create --config <config_file>.yml --port 5001
- Clone the repository
- Install the requirements
- Move to
src/nengo_conversion
folder - Execute the following comand:
nnictl create --config nni_snn_config.yml --port 5001
In test
folder there are notebooks to test the code, specifically Fast-Denser net search, all the NNI experiments performed in this project and then models evaluation for Pytorch+SNNTorch pipeline. Please download it and run it on Google Colab.
NOTE: the results that comes out from Colab are slightly different in terms of accuracy and training time from the ones that come from experiments performed locally with a NVIDIA GTX 1080Ti.
Here are the logs of the NNI experiments. Following that, a summary of the experiments and trials. In the same folder, there are also CSV file generated by Fast-Denser framework to find CNN_K.
- CNN_PT_1 hyperparameters: trial
pOdyC
- CNN_PT_2 hyperparameters: trial
hndeI
- QrbKOuWJ experiments on SNN_PT_2
- 1RSZuxno experiments on SNN_PT_1
- Z19hnw4u experiments on SNN_PT_1
- Best accuracy: trial
U6mOB
- Best tradeoff accuracy/num_steps: trial
yxecs
- Best accuracy: trial
- ZdQn9mbD experiments on SNN_PT_2
- Best accuracy: trial
lcOk8
- Best tradeoff accuracy/num_steps: trial
CQdzk
- Best accuracy: trial
- fykmsgad, NNI experiments on SNN_K:
- Best accuracy: trial
PJTSF
- Best accuracy: trial