amanb2000 / SynRL

Code for the paper "Gradient-Free Neural Network Training via Synaptic-Level Reinforcement Learning"

SynRL: Synaptic Reinforcement Learning

Code for the paper Gradient-Free Neural Network Training via Synaptic-Level Reinforcement Learning. The paper is currently available on arXiv.

@article{bhargava2021gradient,
  title={Gradient-Free Neural Network Training via Synaptic-Level Reinforcement Learning},
  author={Bhargava, Aman and Rezaei, Mohammad R and Lankarany, Milad},
  journal={arXiv preprint arXiv:2105.14383},
  year={2021}
}

Key Files

Novel Work

• Policy file: src/golden_pol2cy.jld2 contains the tabular policy matrix generated in src/04 Multilayer Perceptron.ipynb.
• Simulated decision boundary experiments: src/04 Multilayer Perceptron.ipynb contains code for reproducing the neural network results trained on simulated decision boundaries.
• Synaptic Reinforcement Learing Library: src/SynRLv6.jl is the final set of library functions invoked to train and validate neural networks using the proposed methodology.
• OCR Experiment Script: src/OCR_01.jl is a script that performs an OCR classification experiment using the SynRL library and the hyperparameters passed in through the command line. Results are cached according to command line arguments as well.
• Experiment Orchestration Script: src/orch_02.py is a python script that repeatedly invokes ORC_01.jl with different hyperparameters to run a large set of experiments.

Baseline

• Single Layer Perceptron: src/A1.1 TF Optimized SLP.ipynb contains code used to train and validate the single-layer perceptron using gradient descent on the notMNIST dataset.
• Multilayer Perceptron: src/A2 MLP Gradient Descent.ipynb contains code used to train and validate the multi-layer perceptron using gradiennt descent on the notMNNIST dataset.