self-organizing

Research into self organizing hierarchies for artificial neural networks.

Key highlights

• Accepted at AJCAI 2020. [paper]
• Accepted at NAISys 2020 for poster presentation. [poster]

Installation

To install the dependencies and train, execute the following:

virtualenv -p python3 venv
source venv/bin/activate
chmod +x train.sh 
./train.sh

Note: All arguments and hyperparameters to the program can be modified under the config.yaml file.

Results

• Accuracy vs Max Entropy (global)

• Accuracy vs Meta-steps

TODO

• Refactor the training of handcrafted architectures
• Handle both single and multi-channel inputs (CIFAR10)
• Add code profiler
• Refactor the meta-learner: Simulated Annealing