Why it matters

LTH illuminates network capacity. Understanding shapes sparsity research.

Advertisement

The architecture

Train dense.

Prune by magnitude.

Reset remaining to (early) init.

Retrain: matches dense.

Lottery ticket flowTrain densefullPrune + resetkeep top magnitudeRetrain sparsematches dense acc'Winning ticket' from early training weights works better than pure init
Lottery ticket.
Advertisement

How it works end to end

Iterative magnitude pruning.

Rewind to early training (not init).

Debate: LTH universal or CV-specific.