Grab & Train

Grab the data. Train the model. Learn ML by doing.

Grab the data. Train the model.

Grab & Train is a hands-on machine-learning playground. Draw your own dataset, then train a real neural network right here in your browser and watch it learn the pattern - no code, no setup, no cloud. This is the whole idea of ML in one picture.

🔬 The Lab - train a neural network

idle · draw some points
Class A
Class B
🧽 Erase
Click / drag on the grid to grab data points. Then hit Train. 0 points
Epoch
0
Loss
-
Accuracy
-
Or grab a sample set:
Learning rate0.30
Hidden neurons (per layer)8
Activation
Tip: change a slider mid-training and watch the fit react. Big model + tiny data = overfitting.

What just happened?

Every machine-learning project - from spam filters to image recognition - is really these three steps. You just did all three.

STEP 01

Grab the data

The dots you placed are labelled training data: each has features (its x, y position) and an answer (its colour). A model can only learn patterns that live in the data you give it - garbage in, garbage out.

STEP 02

Train the model

The network starts with random weights and guesses badly. Each epoch it measures how wrong it is (the loss) and nudges every weight a little to do better - that's gradient descent. Watch the loss curve fall as it learns.

STEP 03

Evaluate & tune

The coloured background is the model's decision boundary - its prediction for every point on the grid. High accuracy on your dots is good, but a boundary that wiggles around every single point has likely overfit and won't generalise.

📖 Plain-English glossary

feature - an input the model sees. Here, each point's x and y coordinate.
label - the correct answer for a training example. Here, Class A or Class B.
weights - the numbers inside the network that get adjusted during training. Learning = finding good weights.
epoch - one full pass over all your training data.
loss - a score for how wrong the model is right now. Training tries to push it toward zero.
learning rate - how big a step to take when adjusting weights. Too high overshoots; too low crawls.
activation - the nonlinearity (tanh / ReLU) that lets a network bend into curvy shapes instead of straight lines.
overfitting - memorising the training points instead of learning the general pattern. Watch for a jagged boundary.