āļ„āļđāđˆāļĄāļ·āļ­āļžāļ·āđ‰āļ™āļāļēāļ™

āļŠāļĄāļĄāļ•āļīāļāļēāļ™āļ•āļąāđ‹āļ§āļĨāļ­āļ•āđ€āļ•āļ­āļĢāļĩ

The Lottery Ticket Hypothesis proposes that some large, randomly initialized neural networks contain sparse subnetworks that can train in isolation from their matched initial weights to reach accuracy comparable to the dense model.

āļŸāļąāļ‡āļ„āļģāđāļ™āļ°āļ™āļģāļ™āļĩāđ‰āļšāļĢāļĢāļĒāļēāļĒāđ‚āļ”āļĒ AI · ~2 min

āļšāļ™āļŦāļ™āđ‰āļēāļ™āļĩāđ‰āļ­āđˆāļēāļ™ 2 āļ™āļēāļ—āļĩ
  1. āļ āļēāļžāļĢāļ§āļĄ
  2. āđ€āļˆāļēāļ°āļĨāļķāļ
  3. āļāļēāļĢāđƒāļŠāđ‰āļ‡āļēāļ™āļˆāļĢāļīāļ‡āđƒāļ™āđ‚āļĨāļāđāļŦāđˆāļ‡āļ„āļ§āļēāļĄāđ€āļ›āđ‡āļ™āļˆāļĢāļīāļ‡
  4. āļ­āļ™āļēāļ„āļ•āļ‚āļ­āļ‡āļŠāļĄāļĄāļ•āļīāļāļēāļ™āļ•āļąāđ‹āļ§āļĨāļ­āļ•āđ€āļ•āļ­āļĢāļĩ
  5. āļāļēāļĢāļ™āļģāđ„āļ›āļ›āļāļīāļšāļąāļ•āļī
  6. āļŠāļģāļĢāļ§āļˆāļ•āđˆāļ­āđ„āļ›
  7. āļ„āļģāļ–āļēāļĄāļ—āļĩāđˆāļžāļšāļšāđˆāļ­āļĒ

āļ āļēāļžāļĢāļ§āļĄ

The original experiments support this in tested settings, rather than guaranteeing a compact ticket for every model or task.

āđ€āļˆāļēāļ°āļĨāļķāļ

Jonathan Frankle and Michael Carbin introduced the Lottery Ticket Hypothesis in 2018. In experiments on fully connected and convolutional networks for MNIST and CIFAR-10, they found sparse subnetworks that, when trained in isolation from their original initialization, reached test accuracy comparable to the dense network in a similar number of iterations. Their reported tickets were less than 10–20% of the original model size in those settings; this is an empirical result, not a general compression guarantee. Their iterative magnitude pruning procedure trains a network, removes weights with small magnitudes, restores surviving weights to their initial values, and repeats. Later work studied rewinding to an early training checkpoint. Randomly reinitializing a found mask can reduce the benefit observed in the original experiments. The broader suggestion that dense overparameterization helps optimization discover these sparse subnetworks is a conjecture about the mechanism, not a separate result established by the original experiments.

āļāļēāļĢāđƒāļŠāđ‰āļ‡āļēāļ™āļˆāļĢāļīāļ‡āđƒāļ™āđ‚āļĨāļāđāļŦāđˆāļ‡āļ„āļ§āļēāļĄāđ€āļ›āđ‡āļ™āļˆāļĢāļīāļ‡

Testing whether a sparse subnetwork can retain a dense model’s accuracy on a specified dataset and architecture

Comparing iterative pruning and weight rewinding with training a sparse model from a fresh random initialization

Studying how discovered tickets transfer between related datasets or model architectures

Evaluating whether sparse models reduce storage or inference cost on a target device

āļ­āļ™āļēāļ„āļ•āļ‚āļ­āļ‡āļŠāļĄāļĄāļ•āļīāļāļēāļ™āļ•āļąāđ‹āļ§āļĨāļ­āļ•āđ€āļ•āļ­āļĢāļĩ

Lottery-ticket research asks when sparse subnetworks can be found cheaply, whether they transfer across tasks, and how pruning affects training and inference costs. A separate strong lottery-ticket line studies whether a sufficiently overparameterized random network contains a subnetwork that approximates a target network without further weight training; this is a distinct theoretical claim and not a guarantee about every practical model.

āļāļēāļĢāļ™āļģāđ„āļ›āļ›āļāļīāļšāļąāļ•āļī

āļ—āļģāđ„āļĄāļĄāļąāļ™āļ–āļķāļ‡āļŠāļģāļ„āļąāļ

  • āļāļēāļĢāļ•āļąāļ”āļŠāļīāļ™āđƒāļˆāļ—āļĩāđˆāļŠāļąāļ”āđ€āļˆāļ™āļĒāļīāđˆāļ‡āļ‚āļķāđ‰āļ™. āļŠāđˆāļ§āļĒāđƒāļŦāđ‰āļ„āļļāļ“āđāļĒāļāļ„āļģāļāļĨāđˆāļēāļ§āļ­āđ‰āļēāļ‡āļ—āļēāļ‡āđ€āļ—āļ„āļ™āļīāļ„āļ—āļĩāđˆāļŠāļąāļ”āđ€āļˆāļ™āļ­āļ­āļāļˆāļēāļāļ āļēāļĐāļēāļ—āļēāļ‡āļāļēāļĢāļ•āļĨāļēāļ”āđ„āļ”āđ‰
  • āļ•āđ‰āļ™āļ—āļļāļ™āđāļĨāļ°āļ‡āļšāļ›āļĢāļ°āļĄāļēāļ“. āļ„āļļāļ“āļŠāļēāļĄāļēāļĢāļ–āļ–āļēāļĄāļ„āļģāļ–āļēāļĄāļāļēāļĢāđƒāļŠāđ‰āļ‡āļēāļ™āļ—āļĩāđˆāļ”āļĩāļ‚āļķāđ‰āļ™āļāđˆāļ­āļ™āļ—āļĩāđˆāļˆāļ°āđƒāļŠāđ‰āđ€āļ‡āļīāļ™āļŦāļĢāļ·āļ­āđ€āļ§āļĨāļē
  • āļ—āļĩāļĄāļ‡āļēāļ™āđāļĨāļ°āļ‚āļąāđ‰āļ™āļ•āļ­āļ™āļāļēāļĢāļ—āļģāļ‡āļēāļ™. āļ—āļĩāļĄāļ—āļĩāđˆāļĄāļĩāļ„āļ§āļēāļĄāđ€āļ‚āđ‰āļēāđƒāļˆāļĢāđˆāļ§āļĄāļāļąāļ™āļˆāļ°āļ•āļąāļ”āļŠāļīāļ™āđƒāļˆāđ€āļāļĩāđˆāļĒāļ§āļāļąāļšāļœāļĨāļīāļ•āļ āļąāļ“āļ‘āđŒ āļ™āđ‚āļĒāļšāļēāļĒ āđāļĨāļ°āļāļēāļĢāđ€āļĢāļĩāļĒāļ™āļĢāļđāđ‰āđ„āļ”āđ‰āļ”āļĩāļ‚āļķāđ‰āļ™

āļĢāļ°āļ§āļąāļ‡

  • āđāļ•āđˆāļĨāļ°āļ—āļĩāļĄāļ­āļēāļˆāđƒāļŠāđ‰āļ„āļģāđ€āļ”āļĩāļĒāļ§āļāļąāļ™āļ•āđˆāļēāļ‡āļāļąāļ™ āļ”āļąāļ‡āļ™āļąāđ‰āļ™āļ„āļ§āļĢāļāļģāļŦāļ™āļ”āļ‚āļ­āļšāđ€āļ‚āļ•āļ•āļąāđ‰āļ‡āđāļ•āđˆāđ€āļ™āļīāđˆāļ™āđ†
  • āđ€āļāļ“āļ‘āđŒāļĄāļēāļ•āļĢāļāļēāļ™āļ­āļēāļˆāļ”āļđāđāļ‚āđ‡āļ‡āđāļāļĢāđˆāļ‡āđƒāļ™āļ‚āļ“āļ°āļ—āļĩāđˆāļ›āļĢāļ°āļŠāļīāļ—āļ˜āļīāļ āļēāļžāđƒāļ™āđ‚āļĨāļāđāļŦāđˆāļ‡āļ„āļ§āļēāļĄāđ€āļ›āđ‡āļ™āļˆāļĢāļīāļ‡āđ„āļĄāđˆāđ€āļ—āđˆāļēāļāļąāļ™
  • āļāļēāļĢāđ€āļžāļīāļāđ€āļ‰āļĒāļ•āđˆāļ­āļ„āļļāļ“āļ āļēāļžāļ‚āđ‰āļ­āļĄāļđāļĨāđāļĨāļ°āđāļœāļ™āļāļēāļĢāļ›āļĢāļ°āđ€āļĄāļīāļ™āļĄāļąāļāļŠāļĢāđ‰āļēāļ‡āļœāļĨāļĨāļąāļžāļ˜āđŒāļ—āļĩāđˆāđ€āļ›āļĢāļēāļ°āļšāļēāļ‡

āļ‚āļąāđ‰āļ™āļ•āļ­āļ™āļ•āđˆāļ­āđ„āļ›

  1. āđ€āļĢāļīāđˆāļĄāļ•āđ‰āļ™āļ”āđ‰āļ§āļĒāļ„āļģāļˆāļģāļāļąāļ”āļ„āļ§āļēāļĄāļ āļēāļĐāļēāļ˜āļĢāļĢāļĄāļ”āļēāļ‚āļ­āļ‡āļœāļĨāļĨāļąāļžāļ˜āđŒāļ—āļĩāđˆāļ„āļļāļ“āļ•āđ‰āļ­āļ‡āļāļēāļĢ
  2. āđ€āļĨāļ·āļ­āļāđ€āļĄāļ•āļĢāļīāļāļ§āļąāļ”āļ„āļ§āļēāļĄāļŠāļģāđ€āļĢāđ‡āļˆāļŦāļ™āļķāđˆāļ‡āļĢāļēāļĒāļāļēāļĢāđāļĨāļ°āđ€āļ‡āļ·āđˆāļ­āļ™āđ„āļ‚āļ„āļ§āļēāļĄāļĨāđ‰āļĄāđ€āļŦāļĨāļ§āļŦāļ™āļķāđˆāļ‡āļĢāļēāļĒāļāļēāļĢāļāđˆāļ­āļ™āļāļēāļĢāļ—āļ”āļŠāļ­āļš
  3. āļ”āļģāđ€āļ™āļīāļ™āļāļēāļĢāļ™āļģāļĢāđˆāļ­āļ‡āļ‚āļ™āļēāļ”āđ€āļĨāđ‡āļāļ”āđ‰āļ§āļĒāļ‚āđ‰āļ­āļĄāļđāļĨāļ•āļąāļ§āđāļ—āļ™ āđ„āļĄāđˆāđƒāļŠāđˆāļŠāļļāļ”āļŠāļēāļ˜āļīāļ•āļ—āļĩāđˆāļŠāļ§āļĒāļ‡āļēāļĄ
  4. āđ€āļ­āļāļŠāļēāļĢāļ—āļĩāđˆāļ‹āļķāđˆāļ‡āļŠāļĄāļĄāļ•āļīāļāļēāļ™āđ€āļāļĩāđˆāļĒāļ§āļāļąāļšāļŠāļĨāļēāļāļĨāļ­āļ•āđ€āļ•āļ­āļĢāļĩāđˆāļŠāđˆāļ§āļĒāđ„āļ”āđ‰ āđāļĨāļ°āļ§āļīāļ˜āļĩāļ—āļĩāđˆāļ‡āđˆāļēāļĒāļāļ§āđˆāļēāļˆāļ°āļ”āļĩāļāļ§āđˆāļē

āļŠāļģāļĢāļ§āļˆāļ•āđˆāļ­āđ„āļ›

āļˆāļ”āļŦāļĄāļēāļĒāļ‚āđˆāļēāļ§āļŸāļĢāļĩ

āļĢāļąāļšāļŠāļĢāļļāļ› AI āļĢāļēāļĒāļ§āļąāļ™

āđ€āļĢāļ·āđˆāļ­āļ‡āļĢāļēāļ§ AI āļ—āļĩāđˆāļ•āļĢāļ§āļˆāļŠāļ­āļšāđāļĨāđ‰āļ§ 3 āđ€āļĢāļ·āđˆāļ­āļ‡āļ—āļļāļāļ§āļąāļ™āļ—āļģāļ‡āļēāļ™ āđ€āļŠāđ‰āļē, āđ€āļ‚āļĩāļĒāļ™āđ€āļ›āđ‡āļ™āļ āļēāļĐāļēāļ­āļąāļ‡āļāļĪāļĐāļ˜āļĢāļĢāļĄāļ”āļē āļŸāļĢāļĩāļ•āļĨāļ­āļ”āđ„āļ› āđ„āļĄāđˆāļĄāļĩāđ‚āļ†āļĐāļ“āļē

āļ­āļĩāđ€āļĄāļĨāļŦāļ™āļķāđˆāļ‡āļ‰āļšāļąāļšāļ—āļļāļāļ§āļąāļ™āļ—āļģāļ‡āļēāļ™ āļĒāļāđ€āļĨāļīāļāļāļēāļĢāļŠāļĄāļąāļ„āļĢāđƒāļ™āļ„āļĨāļīāļāđ€āļ”āļĩāļĒāļ§ āđ€āļĢāļēāđ„āļĄāđˆāđ€āļ„āļĒāļ‚āļēāļĒāļŦāļĢāļ·āļ­āđāļŠāļĢāđŒāļ—āļĩāđˆāļ­āļĒāļđāđˆāļ­āļĩāđ€āļĄāļĨāļ‚āļ­āļ‡āļ„āļļāļ“

āļ—āļ”āļŠāļ­āļšāļ•āļąāļ§āđ€āļ­āļ‡

Take the Lottery Ticket Hypothesis quiz

āļ‚āđ‰āļ­āđ€āļŠāļ™āļ­āđāļ™āļ°āļ—āļąāļ™āļ—āļĩāļŠāļģāļŦāļĢāļąāļšāļ—āļļāļāļ„āļģāļ•āļ­āļš āđāļĨāļ°āđƒāļšāļĢāļąāļšāļĢāļ­āļ‡āļ—āļĩāđˆāļŠāļēāļĄāļēāļĢāļ–āđāļŠāļĢāđŒāđ„āļ”āđ‰āļžāļĢāđ‰āļ­āļĄ ID āļ—āļĩāđˆāļ•āļĢāļ§āļˆāļŠāļ­āļšāđ„āļ”āđ‰āđ€āļĄāļ·āđˆāļ­āļ„āļļāļ“āļœāđˆāļēāļ™āļŦāļĨāļąāļāļŠāļđāļ•āļĢ

āđ€āļĢāļīāđˆāļĄāđāļšāļšāļ—āļ”āļŠāļ­āļš

āļŠāļ™āļąāļšāļŠāļ™āļļāļ™āļāļēāļĢāļĻāļķāļāļĐāļē AI āļŸāļĢāļĩ AI Understanding āđ€āļ›āđ‡āļ™āļ­āļ‡āļ„āđŒāļāļĢāđ„āļĄāđˆāđāļŠāļ§āļ‡āļŦāļēāļœāļĨāļāļģāđ„āļĢ 501(c)(3) — āđ„āļĄāđˆāļĄāļĩāđ‚āļ†āļĐāļ“āļē āđ„āļĄāđˆāļĄāļĩāļāļģāđāļžāļ‡āđ€āļ‡āļīāļ™, āļ•āļĨāļ­āļ”āđ„āļ› āļšāļĢāļīāļˆāļēāļ„

āļ„āļģāļ–āļēāļĄāļ—āļĩāđˆāļžāļšāļšāđˆāļ­āļĒ

āļŠāļĄāļĄāļ•āļīāļāļēāļ™āļŠāļĨāļēāļāļĨāļ­āļ•āđ€āļ•āļ­āļĢāļĩāđˆāļ„āļ·āļ­āļ­āļ°āđ„āļĢ?

The Lottery Ticket Hypothesis proposes that some large, randomly initialized neural networks contain sparse subnetworks that can train in isolation from their matched initial weights to reach accuracy comparable to the dense model. The original experiments support this in tested settings, rather than guaranteeing a compact ticket for every model or task.

'āļ•āļąāđ‹āļ§āļ—āļĩāđˆāļŠāļ™āļ°' āđƒāļ™āļŠāļĄāļĄāļ•āļīāļāļēāļ™āļ™āļĩāđ‰āļ„āļ·āļ­āļ­āļ°āđ„āļĢ

āļ•āļąāđ‹āļ§āļ—āļĩāđˆāļŠāļ™āļ°āļ„āļ·āļ­āđ€āļ„āļĢāļ·āļ­āļ‚āđˆāļēāļĒāļĒāđˆāļ­āļĒāļ‚āļ™āļēāļ”āđ€āļĨāđ‡āļāļ‹āļķāđˆāļ‡āđ€āļĄāļ·āđˆāļ­āļāļķāļāđāļĒāļāļˆāļēāļāļ™āđ‰āļģāļŦāļ™āļąāļāđ€āļĢāļīāđˆāļĄāļ•āđ‰āļ™āđ€āļ”āļĩāļĒāļ§āļāļąāļ™ āļˆāļ°āđ„āļ”āđ‰āļ„āļ§āļēāļĄāđāļĄāđˆāļ™āļĒāļģāļ—āļĩāđˆāđ€āļ—āļĩāļĒāļšāđ„āļ”āđ‰āļāļąāļšāđ€āļ„āļĢāļ·āļ­āļ‚āđˆāļēāļĒāļŦāļ™āļēāđāļ™āđˆāļ™

In the original lottery-ticket experiments, what was paired with the discovered sparse mask when retraining a ticket?

The original method paired a discovered pruning mask with the surviving weights reset to their original initialization; later work also studied early-checkpoint rewinding.

āđƒāļ„āļĢāđ€āļ›āđ‡āļ™āļœāļđāđ‰āđ€āļŠāļ™āļ­āļŠāļĄāļĄāļ•āļīāļāļēāļ™āļ•āļąāđ‹āļ§āļĨāļ­āļ•āđ€āļ•āļ­āļĢāļĩ?

Jonathan Frankle āđāļĨāļ° Michael Carbin āđāļ™āļ°āļ™āļģāļŠāļĄāļĄāļ•āļīāļāļēāļ™āđƒāļ™āļĢāļēāļĒāļ‡āļēāļ™āļ‚āļ­āļ‡ MIT āļ›āļĩ 2018

Which procedure did the original Lottery Ticket Hypothesis paper use to identify winning tickets?

The original procedure iteratively pruned low-magnitude weights and reset the surviving weights to their original initialization before retraining; early-checkpoint rewinding was studied later.

How should the idea that overparameterization helps optimization find a trainable sparse structure be described?

The original experiments identify winning tickets in tested settings; the claim that dense overparameterization helps the search for them is a conjecture about why the phenomenon occurs.

āđ€āļĢāļĩāļĒāļ™āļĢāļđāđ‰āļ•āđˆāļ­āđ„āļ›

āļ„āļģāđāļ™āļ°āļ™āļģāļ—āļĩāđˆāđ€āļāļĩāđˆāļĒāļ§āļ‚āđ‰āļ­āļ‡

āļ„āļģāđāļ™āļ°āļ™āļģāđ€āļžāļīāđˆāļĄāđ€āļ•āļīāļĄāļ—āļĩāđˆāđ€āļĨāļ·āļ­āļāļŠāļģāļŦāļĢāļąāļšāļŦāļąāļ§āļ‚āđ‰āļ­āļ™āļĩāđ‰

āļ•āđˆāļ­āđ„āļ›āļ„āļģāđāļ™āļ°āļ™āļģāļ•āđˆāļ­āđ„āļ›

āļĢāļ°āļšāļšāđāļŠāļ”āļ‡āļœāļĨāļ„āļĢāļąāļ§ AI āđāļĨāļ°āļāļģāļŦāļ™āļ”āđ€āļ§āļĨāļēāļ•āļąāđ‹āļ§

āļ­āļļāļ•āļŠāļēāļŦāļāļĢāļĢāļĄ

āļĢāļąāļšāđƒāļšāļĢāļąāļšāļĢāļ­āļ‡ AI āļŸāļĢāļĩāļŦāļĨāļąāļāļŠāļđāļ•āļĢāđāļĨāļ°āļāļēāļĢāļŠāļ­āļšāļŠāļąāđ‰āļ™ āđ† āļŸāļĢāļĩ āđ„āļĄāđˆāļĄāļĩāļ„āđˆāļēāđƒāļŠāđ‰āļˆāđˆāļēāļĒ āđ„āļĄāđˆāļĄāļĩāļšāļąāļ•āļĢāđ€āļ„āļĢāļ”āļīāļ• āļ›āļĢāļ°āļĄāļēāļ“ 2.5 āļŠāļąāđˆāļ§āđ‚āļĄāļ‡āđ€āļĢāļīāđˆāļĄāļ•āđ‰āļ™āļŸāļĢāļĩ →