Pattern Match Learner

A small neural network learns to sort 8×8 black-and-white patterns into six types, in your browser.

Data: 300 generated patterns, 50 of each type. Each has random noise: some pixels flipped on or off.
Network: 64 inputs (one per pixel), one hidden layer with ReLU, 6 outputs with softmax.
Training: stochastic gradient descent on cross-entropy loss, batches of 4 random patterns.
Accuracy: measured on the same 300 patterns it trains on. There is no separate test set.

Press Start learning and watch the loss fall and the weights change. Press Reset network to start again from new random weights. The network usually reaches similar accuracy with different weights each time.

The task

Sort 8×8 pixel patterns into 6 types. 300 training patterns in total. Examples without noise:

H-STRIPE

V-STRIPE

DIAGONAL

CHECKER

X-CROSS

FRAME

All network weights

Lighter = higher value. Orange = large gradient on the last step.

Network activations

INPUT (64) HIDDEN (6) OUTPUT (6)

Node brightness = activation. Line strength = weight size.

Try the network

Click cells to draw an 8×8 pattern. The network classifies it as you draw.

Prediction: —

Confidence: —

Hidden layer (6 units)

Output layer (6 classes)

for each batch:
  gradient = ∂Loss/∂θ
  θ = θ - α · gradient

where:
  θ = network parameters (weights)
  α = learning rate (step size)

Learning rate (α)

α = 0.100

Hidden units (resets network)

Speed (1.0 = 100 steps per second)

0.2x

Max iterations (0 = no limit)

Training stops at this step count. 0 means it runs until you pause it.

Iteration: 0

Loss: 0.000

Accuracy: 0%

Status: Ready