Backpropagating with matrices (instead of raw scalars) can help make the task much easier. My neural network looks like:

Input Layer Nodes: $I_a (Ate Cookies?)$, $I_b (Drank Milk?)$

Hidden Layer 1 Nodes: $H_c$, $H_d$, $H_e$

Hidden Layer 2 Nodes: $H_f$, $H_g$, $H_h$

Output Layer Nodes: $O_i (Full)$, $O_j (Unhappy)$

Expected Output: Drinking Milk AND Eating Cookies = Full, Drinking Milk XOR Eating Cookies = Unhappy, Neither = Nothing

Activation Function: $ActivateFunc(N_x)=\frac{1}{1+e^{-N_x}}$

Actual Output: (Cookies 1, Milk 1) -> (Unhappy 1, Full 1)

Weight Notation: $ac$ represents a connection from $I_a$ to $H_c$, for example.

I know that the weight update formula is: $w(n+1)=w(n)+R_{learning} \cdot w(n-1) + \eta \cdot \delta(n) \cdot y$

All I know so far about neural network matrices is that there is a matrix for the input layers, output layers, hidden layers, and weights.

How can I apply backpropagation to the neural network in matrix form?


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.