🔄 3D Continual Learning: Elastic Weight Consolidation

The weight vector w = (w1, w2, w3) starts at the origin. Each of the four tasks has a fixed target weight vector w*_k and a per-axis Fisher-information importance F_k, and its loss is the real quadratic form used by EWC's own approximation: L_k(w) = ½·Σ F_k,i·(w_i − w*_k,i)². Pressing "Train" runs genuine gradient descent on the current task's loss; when EWC protection is on, the gradient also includes λ·Σ_{j<k} F_j·(w − w*_j) — the actual elastic penalty from Kirkpatrick et al.'s Elastic Weight Consolidation, pulling the point back toward every earlier task's optimum in proportion to how important each of its weights was. Turn EWC off and train several tasks to watch the weight vector drift freely to the newest optimum, tanking earlier tasks' accuracy — catastrophic forgetting, reproduced with the same numbers that drive the visualization.

3D · Three.js / WebGL renderer · 60 FPS target · runs fully client-side, no install