**The Math Behind Backpropagation**

Published 02/27/2016

==================================

## ## Introduction

Previously we wrote a short introduction to neural networks, which discusses backpropagation as the training method of choice, described as: "simply a way of minimizing the loss function, or error, of the network by propagating errors backward through the network and adjusting weights accordingly."

This brief writeup is meant to shed light on the mathematics behind backpropagation, deriving (with substantial justification) the weight changing algorithm for a feedforward neural network by means of a standard gradient descent.

## ## The feedforward algorithm

The activation of a neural network is iteratively defined by

$\begin{aligned} y_n &= f(x_n) \\ x_{n+1} &= w_n y_n \end{aligned}$where $y_n$ is the output vector at layer $n$, $f$ is the activation function, $x_n$ is the input vector at layer $n$, and $w_n$ is the weight matrix between layers $n$ and $n+1$. The first layer is $n=1$ and the last layer is $n=N$.

## ## The backpropagation algorithm

The error of the network is defined by

$\begin{aligned} c &= \frac{1}{2}(y_N - t)^2 \end{aligned}$The error gradient of the input vector at a layer $n$ is defined as

$\begin{aligned} \delta_n = \frac{\partial c}{\partial x_n} \end{aligned}$The error gradient of the input vector at the last layer $N$ is

$\begin{aligned} \delta_N &= \frac{\partial c}{\partial x_N} \\ &= \frac{\partial}{\partial x_N} \frac{1}{2}(y_N - t)^2 \\ &= \left( \frac{\partial}{\partial y_N} \frac{1}{2}(y_N - t)^2 \right) \frac{\partial y_N}{\partial x_N} \\ &= (y_N - t) \frac{\partial f(x_N)}{\partial x_N} \\ &= (y_N - t) f'(x_N) \end{aligned}$The error gradient of the input vector at an inner layer $n$ is

$\begin{aligned} \delta_n &= \frac{\partial c}{\partial x_n} \\ &= \frac{\partial c}{\partial x_{n+1}} \frac{\partial x_{n+1}}{\partial x_n} \\ &= \delta_{n+1} \frac{\partial x_{n+1}}{\partial x_n} \\ &= \delta_{n+1} \frac{\partial w_n y_n}{\partial x_n} \\ &= \delta_{n+1} \frac{\partial w_n y_n}{\partial y_n} \frac{\partial y_n}{\partial x_n} \\ &= \delta_{n+1} \frac{\partial w_n y_n}{\partial y_n} \frac{\partial f(x_n)}{\partial x_n} \\ &= \delta_{n+1} w_n f'(x_n) \end{aligned}$Therefore, the error gradient of the input vector at a layer $n$ is

$\begin{aligned} \delta_n &= f'(x_n) \begin{cases} (y_N - t) & \text{if } n = N \\ \delta_{n+1} w_n & \text{if } n < N \end{cases} \end{aligned}$Hence, the error gradient of the weight matrix $w_n$ is

$\begin{aligned} \frac{\partial c}{\partial w_n} &= \frac{\partial c}{\partial x_{n+1}} \frac{\partial x_{n+1}}{\partial w_n} \\ &= \delta_{n+1} \frac{\partial w_n y_n}{w_n} \\ &= \delta_{n+1} y_n \end{aligned}$Therefore, the change in weight should be

$\begin{aligned} \Delta w_n &= -\alpha \frac{\partial c}{\partial w_n} \\ &= -\alpha \delta_{n+1} y_n \end{aligned}$where $\alpha$ is the learning rate (or rate of gradient descent). Thus, we have shown the necessary weight change, from which the implementation of a training algorithm follows trivially.

## ## Citation

Cited as:

Martin, Carlos G. and Schuermann, Lucas V. (Feb 2016).

The Math Behind Backpropagation. Writing. https://lucasschuermann.com/writing/math-behind-backpropagation

Or

```
@misc{martin2016backpropagation,
title = {The Math Behind Backpropagation},
author = {Martin, Carlos G. and Schuermann, Lucas V.},
year = {2016},
month = {Feb},
url = "https://lucasschuermann.com/writing/math-behind-backpropagation"
}
```