An implementation of predictive coding in Python, utilising pytorch for GPU acceleration. We compare the use of the free energy as a loss function vs mean squared error. In addition we also test these predictive coding networks against a normal backpropagation network.

Running

Free Energy Loss Function

python -m scripts.supervised

Mean Squared Error Loss Function

python -m scripts.supervised_mse

Backpropagation

python -m scripts.supervised_bp

Requirements

• numpy
• torch
• torchvision

Authors

• Alexander Tschantz
• Beren Millidge

Extended by

• Jose Urruticoechea
• Panos_Syrgkanis
• Serdar Sungun