Source
Evci, Ioannou, Keskin and Dauphin (2020), LIT-039.
What was actually shown
The question is the one THEORY-002 leaves open: why does a sparse structure trained from a random initialization do worse than the same structure obtained by pruning a dense network? The answer offered is gradient flow — sparse networks at initialization have poor flow, and training from scratch has to overcome that before it can make progress on the task.
Both halves are LIT-039’s: it attributes the from-scratch gap to poor gradient flow in sparse networks at initialization, and finds that a rewound ticket converges back to the solution of the dense run it was pruned from.
The part that bears on the ticket is the deflation. A rewound ticket is followed and found to re-learn the solution its dense pruning run had already reached: it lands in the same basin rather than discovering, independently, that this sparse structure was a good one.
What it corrects
Not the existence result. The measurement in THEORY-002 stands, and nothing here says the ticket fails to match the dense network.
What breaks is the reading — that a search has located a sparse architecture worth having. The claim that survives is narrower and more specific: the initialization of the surviving weights carries information about the solution the dense run arrived at, and the ticket’s job is to get back there. “Sparse architectures are findable” becomes “the initialization remembers”, and only the second is supported.
What this does not say
It does not say the phenomenon is uninteresting. That initial values carry recoverable information about a solution reached much later is a strong statement about training dynamics, and it is the reason both of these documents are in the record.
It does not license a recommendation, in either direction. The record holds no pruning or sparsity practice, and neither of these two documents is the thing that would create one — which is the state this scheme was added to be able to file honestly rather than as a gap in the practice registry.