Probabalistic Learning Framework implementing surpervised learning to solve the MNIST hand-written digit recognition problem using Neuromorphic Spiking Neural Networks. Performance of the models were also evaluated when transferring weights from floating point to low precision weights/using Phase Change Memory such as that would be in the case of transferring the model to a Neuromorphic based chip.
Several learning rate functions were experimented with:
4 bit precision with a standard deviation of 0.1
- The exponentially decaying learning rate was the best performer.
- Cyclic learning rates help if time is a contraint!
