A Python implementation of An Approximation of the Error Backpropagation Algorithm in a Predictive Coding Network with Local Hebbian Synaptic Plasticity
[Paper]
Based on the MATLAB implementation from [@djcrw]
numpytorchtorchvision
- Include model from A tutorial on the free-energy framework for modelling perception and learning
- Add additional optimisers
- Measure number of iterations
- The initial space of mu needs to be sufficently large - ensembles of amortised weights or slow learning rate?
- Test pure PC accuracy
- Errors go down, but amortised asymptotes
- Infinite iterations - amortised learning provides mechanism for setting number, remove free parameter (replaces with threshold)
- Generative overtakes and this is inconsistent with discriminative - need some way to promote discrimination in network
- Both trying to predict each other? Generative predicting discrimantive?
- Add gradient clamping