Perceptron Mistake Bound: Margin, Radius and Weight Vector Convergence (2D)
The same perceptron model as the 3D scene: a top-down plane of two classes, the true separator with its margin band, the learning perceptron weight vector and separator, and the cosine to the true normal after each update.
Train a perceptron on linearly separable points and watch its weight vector turn toward the true separator after every mistake. The Novikoff theorem says the number of mistakes can never exceed (R/gamma)^2, where R is the data radius and gamma the margin. Change the margin, radius, number of points and separator angle, with checks of the proof inequalities, rotation and scaling invariance built in. A Blender-built pegboard.
The same perceptron model as the 3D scene: a top-down plane of two classes, the true separator with its margin band, the learning perceptron weight vector and separator, and the cosine to the true normal after each update.
2D · HTML5 Canvas 2D · 60 FPS target · runs fully client-side, no install