Forward propagation
End-to-end, input-to-output transformation in a neural network:
\begin{equation*} \hat{Y} = \hat{f}(X) = A^{[\ell]} \in \mathbb{R}^{p^{\scriptstyle [\ell]} \times p^{\scriptstyle [\ell - 1]}} \end{equation*}with
\begin{equation*} A^{[i]} = \begin{cases} X \in \mathbb{R}^{n \times m} & i = 0 \\ \boxed{g^{[i]}(W^{[i]} A^{[i - 1]} + \vec{b}^{[i]})} \in \mathbb{R}^{p^{\scriptstyle [i]} \times p^{\scriptstyle [i - 1]}} & i > 0 \end{cases} \end{equation*}and
\begin{equation*} p^{[1]} = p^{[0]} \end{equation*}where
- \(X_{m \times n}\) is the input matrix
- \(g^{[i]}\) is the activation function used in the \(i\)-th layer
- \(W^{[i]}\) is a matrix in the weight tensor
- \(\vec{b}^{[i]}\) is a vector in the bias tensor
- \(p^{[i]}\) is the number of neurons in the \(i\)-th layer
- \(\ell\) is the number of layers in the network.
- Dimensionality check: Base case
- Dimensionality check: Recursive case
- Synonyms