Robust-Classification-EBN

Prerequisites

In order to run the scripts, the following packages must be installed

• rockpool using the feature/NetworkADS-Jax and the feature/FORCE-Jax branch

• SIMMBA using the fix/data-use-only branch (not required for temporal XOR task)

• JAX (version 1.0.75)

See requirements.txt for some packages needed.

General Remarks

All the models have been pre-trained and are contained in this repository. If you would like to re-train any model, you need to adjust the paths used in the scripts and delete/move the pre-trained model file, which is usually contained in a folder named Resources.

The implementation of the learning rule can be found in rockpool/rockpool/layers/gpl/lif_jax.py. There you will find the implementation of the ADS network given that you have checked out feature/NetworkADS-Jax from the repository mentioned above. This file also contains the FORCE implementation, which follows the original MATLAB implementation.

Figure 1

To reproduce figure 1, navigate to the folder figure1 and execute

$ python figure1.py

This will execute the script and train the model. Afterwards, you can generate the data necessary for plotting by executing

$ python generate_data.py

This will execute the script and save all the data necessary for plotting to a folder named Plotting. Afterwards one can execute $ python plot.py to generate the figure.

NOTE:

• You have to adjust the paths used in both files.
• By deleting Resources/temporal_xor.json and then running the script figure1.py, you will retrain the model.

Figure 2

Execute $ python generate_data.py and $ python plot.py to generate the data needed for plotting and to create Figure 2.

NOTE:

• We use a pre-trained model in mismatch/Resources. If you want to re-train this network, go to mismatch/NetworkADS and train the model using the --use-ebn and --network-idx 0 flag.
• Data availability: The model is being trained and validated on data that is not directly available to the community. The core wake-phrase data can be obtained from here. The data was augmented with some background noise. If you have filled out the form on the previous link, please contact me for the entire data. This is for research purposes only.
• After you have the data HeySnips and DEMAND, navigate to path/to/simmba/simmba/SIMMBA/experiments/ and open HeySnipsDEMAND.py where you need to adjust the paths.

Robustness

Discretization

The folder discretization/ contains one sub-folder for each network architecture. In each subfolder, there is a script to analyze each network instance for different levels discretization of the recurrent weights. The levels of weight-precision are: 4,5, and 6 bits.

Figure 3

Navigate to the folder figure3/ and execute python generate_data.py, which will generate and store the data needed for the figure.

Afterwards execute python plot.py to generate the figure.

Sudden Neuron Death

In the folder suddenNeuronDeath/ execute bash compare_networks.sh. This will iteratively execute compare_jax.py for each trained network instance. If you want to analyze only one instance, execute python compare_jax.py --network-idx X.

Afterwards, execute python compute_results.py to compute and display the results of the analysis.

Component Mismatch Analysis with ADS Network, FORCE, Reservoir and BPTT

In this section, the robustness to component mismatch is investigated for different network architectures/learning schemes. Please navigate to the folder path/to/this/repo/mismatch.

The folder contains sub-folders for each architecture. Each sub-folder contains a script for training a network and analyzing it. If you would like to train a network, for example for NetworkADS, delete or copy the existent file from Resources/ to some other place and execute python hey_snips_ads_network_jax.py --network-idx X where X is any number you can choose.

The mismatch analysis is carried out for each network instance to ensure statistical correctness and can be carried out by executing bash mismatch_analysis.sh for each architecture.

Injected Noise

In this experiment we analyze the robustness to noise injected directly into the membrane potentials for different architectures.

Navigate to the folder membranePotentialNoise/<architecture> and execute bash noise_analysis.sh to analyze all network instances.