LearnDeep learning
Perceptron
Meet the first learning machine: one artificial neuron that corrects its line one mistake at a time.
Press Step to fix one mistake, or Run to keep going. Circled points are on the wrong side. Click to add points in the chosen colour.
How it works
One neuron, one line.
The perceptron, invented by Frank Rosenblatt in 1958, is the ancestor of every neural network. It weighs its inputs, adds a bias and answers pink if the total is above zero: ŷ = sign(w₁x + w₂y + b). Geometrically, that is a straight line with pink on one side and blue on the other.
It learns from mistakes only. Whenever a point lands on the wrong side, it adds that point to its weights or subtracts it, which turns and shifts the line towards the point. If some straight line can separate the two colours, this simple rule is guaranteed to find one after a finite number of mistakes.
If no straight line can separate them, as with XOR, it never settles. Showing this in 1969 slowed neural network research for years. The fix is to stack neurons in layers, which is the next simulation.