The lottery ticket hypothesis: why neural networks work

Breaking centuries of accepted theory, today's most powerful AI systems succeed by defying what was once considered mathematical law. Five years ago, training neural networks with trillions of parameters would have been dismissed as foolish.

"Bigger models just overfit," was the mantra, backed by the bias-variance tradeoff principle that had governed learning systems for over 300 years. The logic seemed unassailable: make your model too simple, it misses patterns; make it too complex, it memorizes noise instead of signals.

Then in 2019, researchers committed scientific heresy - they scaled neural networks far beyond the point where theory predicted catastrophic failure. Instead of collapsing, these massive models showed "double descent," where after initially overfitting, performance dramatically improved again.

The explanation came from MIT's "lottery ticket hypothesis": large networks succeed not by memorizing, but by containing countless potential simple solutions with different starting conditions. Training becomes a massive lottery draw where the best-initialized small network emerges victorious.

This revelation reconciles empirical success with classical theory. Intelligence isn't about memorizing complexity - it's about finding elegant patterns that explain complex phenomena. Scale simply provides more lottery tickets, more chances to find those optimal solutions.