SGD Classifier Visualizer

Select a class and click to place points

Specify data

Type

Configure model

Loss

Learning rate

Batch size

Mini-batch selection

Each point is used once before the dataset is shuffled again.

Additional options

Random seed

Animation speed

Metrics

Loss Accuracy

Parameters

Step 0 of 0

Class A parametersWeight vector and bias for the positive class

w_A = —
b_A = —

Pending Class A updateNegative learning rate × derivatives

Δw_A = —
Δb_A = —

Add at least one point from each class to begin.

Example scenarios

Load a scenario, then use Next or Animate in the Train tab to observe its behavior.

This uses the easy linearly separable dataset and the perceptron loss. The decision boundary eventually converges to one with zero loss and perfect classification accuracy on the data.

This uses the perceptron loss and hard linearly separable dataset. The decision boundary eventually converges to one with zero loss and perfect classification accuracy on the data.

This uses the perceptron loss and the non-linearly separable dataset. The decision boundary does not converge.

This uses squared error and the hard linearly separable dataset with the default learning rate of 1. The parameters diverge and the decision boundary does not converge.

This uses squared error and the hard linearly separable dataset with a small learning rate of 0.1. The parameters and decision boundary eventually fluctuates slightly around a perfect classifier.

This uses log loss and the hard linearly separable dataset. The parameters and decision boundary eventually fluctuates very slightly around a perfect classifer.

Using the visualizer

Choose data. Select a built-in dataset, or choose Custom data and click the plot to add Class A and Class B points.
Configure SGD. Select a loss, learning rate, mini-batch size, and selection strategy. A fixed random seed reproduces the same batch order.
Open Train. Points with amber rings form the mini-batch that will be used by the next update.
Step through training. Next performs one SGD step. Previous and Reset inspect earlier states without deleting later history.
Animate. Animate advances automatically at the configured speed; Pause stops it.

Reading the display

The solid line is the current decision boundary. The dotted line is the initial reference boundary.
Background colors show the predicted class regions.
The Parameters panel shows Class A's current weight and bias and its pending update.
The Metrics chart evaluates loss and accuracy on the entire dataset. Larger outlined markers identify the step currently displayed.
Hover over any point to inspect its class and coordinates.

Decision boundary