appu226 / paragraph Goto Github PK

As a user,
I want the traditional backward propagation algorithm to be implemented,
So that gradient computation has smaller peak memory requirement.

The current algorithm computes the values and derivatives in a single pass from inputs to outputs.
Instead, implementing the traditional backpropagation algorithm, which first computes the values from inputs to outputs, and them computes the gradients from the outputs to the inputs, would be more efficient.

Use cmake instead of Makefile for more portability.

As a user of this library,
I want a perceptron training algorithm to be implemented,
So that I can:

• train perceptrons
• have an example for building other training algorithms

Currently the softmax node normalises the sigmoid probabilities across all dimensions. This would mean, per tensor, we are classifying a single point.
Instead, it should normalise across only one dimension, so that per tensor we can classify multiple points.

As a user,
I want the tensors to be implemented in a sparse manner,
So that peak memory requirement is reduced.