Neural Network Overfitting & Regularization Lab

Live Backprop Runtime
2D Spatial Decision Boundary
Noisy Sports Regime Circles Two Spirals Moons
Click Mode:
Class A (Purple)
Class B (Green)
Erase
Interactive 2D Space: Click or drag inside the canvas to add or remove ground-truth training samples. Watch decision contours flex during backpropagation.
Model Convergence & Telemetry
Epochs 0
Train Loss 0.0820
Val Loss 0.1140
Gen Gap 0.0320
Boundary Smoothness: Convex Regularized
Model parameters are constrained by L2 penalty weight decay (0.0050) and Dropout (0.10), curbing sharp overfitting boundaries and controlling the generalization gap.
Hyperparameters & Regularization
Network Capacity (Hidden Layers)
Learning Rate (η) 0.030
L2 Regularization (λ) 0.0050
Dropout Probability (p) 0.10
Dataset Noise / Variance 0.15
Hidden Activation
Overfitting Trap: Increasing hidden units without regularization allows the model to memorize noisy training points (high variance, zero train loss, huge val loss gap). L2 weight decay smoothly pulls boundaries convex.
Enjoy this tool? Build your own with Super