Source
Frankle and Carbin (2018), LIT-019.
What was actually shown
Iterative magnitude pruning, then a rewind. Train the dense network, remove the smallest-magnitude weights, restore the survivors to the values they were initialized with, and train that sparse network alone. It reaches the dense network’s accuracy in at most the same number of iterations, at 10–20% of the parameters, on the paper’s vision benchmarks — sometimes faster and sometimes generalizing better.
The control is what makes it a finding rather than a pruning result. Keep the same sparse structure and re-initialize it randomly, and it does not work. So what the dense run produced is not a better architecture; it is a pairing of a structure with particular initial values, and neither half suffices.
Why this is Proposed rather than Active
The result reproduces and is not in dispute on the networks it was run on. What is unsettled is the scope, and the evidence against generality has a specific shape: at larger scale, rewinding to initialization stopped working and later work rewound to an early-training checkpoint instead. That is a weaker claim wearing the same name — if the information has to be collected over the first k steps, it was not in the initialization.
THEORY-004 is a second reason for the caution, and it is about what the result means rather than where it holds.
What this does not say
It does not say you can find the ticket without the dense run. The procedure needs a full dense training pass before it produces anything, which is why LIT-019 carries no practice: as an instruction it reads “train the model, then train a smaller one.”
It does not say sparse structures are independently good architectures. That is the reading the phrase “winning ticket” invites, it is the one THEORY-004 takes apart, and the random-reinitialization control in this very paper already argues against it.