howardwang15 / neural-net-scratch Goto Github PK

The input layer should be a placeholder for incoming data. Should include fields such as

• name
• shape (excluding batch size)

Add a base layer that all other layers inherit from. Attributes of this layer include:

• name

Base layer should also define interfaces (functions) other layers have to implement

• forward
• backprop

Building a neural network library from scratch will require rigorous testing early on. Explore different testing frameworks and create a simple test for the base layer

Most basic type of layer that performs computation is the dense layer. This layer should create

• weights
• biases

The shape of the weights and biases depend on the previous layer's output shape and the number of dimensions the dense layer should output.

Should also include

• activation function
• weight initialization methods
• regularization