ML//neural network//perceptron

The perceptron is the original single-neuron classifier (Rosenblatt, 1958): it weights each input, sums them, adds a bias and thresholds the result.

The perceptron is the original single-neuron classifier (Rosenblatt, 1958): it weights each input, sums them, adds a bias and thresholds the result.

It computes z=w⋅x+bz = \mathbf{w} \cdot \mathbf{x} + bz=w⋅x+b, then fires 1 if zzz clears a threshold, else 0. Geometrically it draws a single hyperplane: one side says yes, the other no.

It learns by nudging its weights toward a less wrong answer after each mistake, the seed of all gradient descent training.

Its ceiling is hard: one hyperplane only separates linearly separable data (it cannot learn XOR). The fix is stacking neurons into a neural network.