Linear Regression and Adaptive Linear Neurons (Adalines) are closely related to each other. In fact, the Adaline algorithm is a identical to linear regression except for a threshold function $\phi(\cdot)_T$ that converts the continuous output into a categorical class label

where $z$ is the net input, which is computed as the sum of the input features $\mathbf{x}$ multiplied by the model weights $\mathbf{w}$:

(Note that $x_0$ refers to the bias unit so that $x_0=1$.)

In the case of linear regression and Adaline, the activation function $\phi(\cdot)_A$ is simply the identity function so that $\phi(z)_A = z$. Now, in order to learn the optimal model weights $\mathbf{w}$, we need to define a cost function that we can optimize. Here, our cost function $J({\cdot})$ is the sum of squared errors (SSE), which we multiply by $\frac{1}{2}$ to make the derivation easier:

where $y^{(i)}$ is the label or target label of the $i$th training point $x^{(i)}$.

(Note that the SSE cost function is convex and therefore differentiable.)

In simple words, we can summarize the gradient descent learning as follows:

1. Initialize the weights to 0 or small random numbers.
2. For $k$ epochs (passes over the training set)
1. For each training sample $x^{(i)}$
1. Compute the predicted output value $\hat{y}^{(i)}$
2. Compare $\hat{y}^{(i)}$ to the actual output $y^{(i)}$ and Compute the "weight update" value
3. Update the "weight update" value
2. Update the weight coefficients by the accumulated "weight update" values

Which we can translate into a more mathematical notation:

1. Initialize the weights to 0 or small random numbers.
2. For $k$ epochs

1. For each training sample $x^{(i)}$

1. $\phi(z^{(i)})_A = \hat{y}^{(i)}$
2. $\Delta w_{(t+1), \; j} = \eta (y^{(i)} - \hat{y}^{(i)}) x_{j}^{(i)}\;$ (where $\eta$ is the learning rate);
3. $\Delta w_{j} := \Delta w_j\; + \Delta w_{(t+1), \;j}$
2. $\mathbf{w} := \mathbf{w} + \Delta \mathbf{w}$

Performing this global weight update

can be understood as "updating the model weights by taking an opposite step towards the cost gradient scaled by the learning rate $\eta$"

where the partial derivative with respect to each $w_j$ can be written as

To summarize: in order to use gradient descent to learn the model coefficients, we simply update the weights $\mathbf{w}$ by taking a step into the opposite direction of the gradient for each pass over the training set -- that's basically it. But how do we get to the equation

Let's walk through the derivation step by step.