An interactive demo of how neural networks learn through error calculation, gradients, and weight updates.
This demo intentionally uses a small neural network so the essential learning mechanisms remain visible and understandable.
Data
Model
Speed
Didactics
Underfitting: model too simple, high bias, poor training and test performance. Overfitting: model too flexible, low training error but weaker generalization.
Visualization
Das langfristig gelernte Wissen eines neuronalen Netzes steckt nicht in einem einzelnen Neuron, sondern verteilt in Gewichten und Biases. Dickere Verbindungen zeigen stärkere gelernte Kopplungen.
y_hat = f(x, W, b)L = -[y * log(y_hat) + (1 - y) * log(1 - y_hat)]gradient = dL / dWW_new = W_old - learning_rate * gradientGradient descent moves in the direction opposite to the loss gradient.Precision: 0.0% · Recall: 0.0% · F1: 0.0%
Training and test performance are compared after the first training steps.
Start training or the guided explanation. The network begins with random weights and learns a decision function.
This demo shows a small neural network with two inputs, hidden layers, and one output. Each point has two coordinates and belongs to one of two classes. The weights start randomly, so predictions are initially unreliable. In the forward pass, the network calculates a probability for class 1. Loss measures how far that prediction is from the true class. Backpropagation uses the chain rule to calculate derivatives for all weights and biases. Gradient descent then updates parameters in small steps to reduce loss. Accuracy typically rises and the decision boundary adapts to the data structure. The network learns a parameterized function, not explicit rules.