Π Π£ΠšΠžΠ’ΠžΠ”Π‘Π’Π’Πž ПО ΠžΠ‘ΠΠžΠ’ΠΠœ

Π“Ρ€ΠΎΠΊΠΊΠΈΠ½Π³ ΠΈ отсрочСнная гСнСрализация

Grokking describes studied cases where held-out generalization improves well after a network has fit its training examples.

ΠŸΡ€ΠΎΡΠ»ΡƒΡˆΠ°ΠΉΡ‚Π΅ это Ρ€ΡƒΠΊΠΎΠ²ΠΎΠ΄ΡΡ‚Π²ΠΎΠžΠ·Π²ΡƒΡ‡Π΅Π½ΠΎ ИИ Β· ~3 min

На этой страницС3 ΠΌΠΈΠ½ΡƒΡ‚Ρ‹ чтСния
  1. ΠžΠ±Π·ΠΎΡ€
  2. Π“Π»ΡƒΠ±ΠΎΠΊΠΎΠ΅ ΠΏΠΎΠ³Ρ€ΡƒΠΆΠ΅Π½ΠΈΠ΅
  3. РСальная рСализация
  4. Π‘ΡƒΠ΄ΡƒΡ‰Π΅Π΅ Π³Ρ€ΠΎΠΊΠΊΠΈΠ½Π³Π° ΠΈ отсрочСнной Π³Π΅Π½Π΅Ρ€Π°Π»ΠΈΠ·Π°Ρ†ΠΈΠΈ
  5. ΠŸΡ€ΠΈΠΌΠ΅Π½Π΅Π½ΠΈΠ΅ Π½Π° ΠΏΡ€Π°ΠΊΡ‚ΠΈΠΊΠ΅
  6. ΠŸΡ€ΠΎΠ΄ΠΎΠ»ΠΆΠ°ΠΉΡ‚Π΅ ΠΈΡΡΠ»Π΅Π΄ΠΎΠ²Π°Ρ‚ΡŒ
  7. Часто Π·Π°Π΄Π°Π²Π°Π΅ΠΌΡ‹Π΅ вопросы

ΠžΠ±Π·ΠΎΡ€

Small algorithmic tasks provide clear experiments, but neither a validation plateau nor longer training guarantees that this transition will occur.

Π“Π»ΡƒΠ±ΠΎΠΊΠΎΠ΅ ΠΏΠΎΠ³Ρ€ΡƒΠΆΠ΅Π½ΠΈΠ΅

The original grokking paper, publicly submitted in January 2022, studies small algorithmically generated datasets. In some experiments, a network fits the training data while held-out performance remains poor, then generalizes after substantially more optimization. This separates fitting observed examples from learning behavior that transfers to unseen examples of the underlying task. It does not mean that every model has near-zero validation accuracy during an initial plateau or that memorization must always disappear completely. The original study varies dataset sizes and optimization choices, including weight decay. Its findings concern those setups rather than a unique universal explanation of generalization. Progress Measures for Grokking via Mechanistic Interpretability investigates small transformers trained on modular addition. The studied algorithm uses Fourier structure and trigonometric identities to represent addition through circular relationships. The researchers examine activations and weights and use Fourier-space ablations to test their account. They describe memorization, circuit formation, and cleanup phases, with generalizing structure emerging before the visible jump in test accuracy. Weight decay removes memorization components during the studied cleanup process. This is more specific than a story in which the entire algorithm appears suddenly at the accuracy jump. Modular tasks offer exact labels: for example, adding six and five modulo seven gives four because eleven leaves remainder four after division by seven. The original paper also reports increasing optimization requirements as training datasets become smaller. Such controlled experiments motivate further research, but do not establish that a large deployed model will eventually improve if trained indefinitely.

РСальная рСализация

Plotting training and held-out accuracy on a modular-arithmetic experiment.

Comparing regularization settings on a defined small algorithmic dataset.

Inspecting the Fourier-based circuit studied in a small transformer.

Varying training-set size while recording the optimization needed for held-out performance.

Π‘ΡƒΠ΄ΡƒΡ‰Π΅Π΅ Π³Ρ€ΠΎΠΊΠΊΠΈΠ½Π³Π° ΠΈ отсрочСнной Π³Π΅Π½Π΅Ρ€Π°Π»ΠΈΠ·Π°Ρ†ΠΈΠΈ

Controlled experiments can test how dataset size, architecture, regularization, and optimization affect delayed generalization. Progress measures based on a verified circuit may reveal behavior hidden by an overall accuracy curve, though scaling such analysis to larger models remains an open question. Researchers should distinguish a measured mechanism in a small task from a proposed explanation for every model. Practical stopping decisions still need evaluation, resource constraints, and evidence of progress; the existence of grokking provides a reason to investigate some plateaus rather than a rule to continue all training indefinitely.

ΠŸΡ€ΠΈΠΌΠ΅Π½Π΅Π½ΠΈΠ΅ Π½Π° ΠΏΡ€Π°ΠΊΡ‚ΠΈΠΊΠ΅

ΠŸΠΎΡ‡Π΅ΠΌΡƒ это Π²Π°ΠΆΠ½ΠΎ

  • Π‘ΠΎΠ»Π΅Π΅ Ρ‡Π΅Ρ‚ΠΊΠΈΠ΅ Ρ€Π΅ΡˆΠ΅Π½ΠΈΡ. Π­Ρ‚ΠΎ ΠΏΠΎΠΌΠΎΠΆΠ΅Ρ‚ Π²Π°ΠΌ ΠΎΡ‚Π΄Π΅Π»ΠΈΡ‚ΡŒ Ρ‡Π΅Ρ‚ΠΊΠΈΠ΅ тСхничСскиС заявлСния ΠΎΡ‚ ΠΌΠ°Ρ€ΠΊΠ΅Ρ‚ΠΈΠ½Π³ΠΎΠ²ΠΎΠ³ΠΎ языка.
  • Π‘Ρ‚ΠΎΠΈΠΌΠΎΡΡ‚ΡŒ ΠΈ Π±ΡŽΠ΄ΠΆΠ΅Ρ‚. Π’Ρ‹ ΠΌΠΎΠΆΠ΅Ρ‚Π΅ Π·Π°Π΄Π°Ρ‚ΡŒ Π±ΠΎΠ»Π΅Π΅ эффСктивныС вопросы ΠΏΠΎ Ρ€Π΅Π°Π»ΠΈΠ·Π°Ρ†ΠΈΠΈ, ΠΏΡ€Π΅ΠΆΠ΄Π΅ Ρ‡Π΅ΠΌ Ρ‚Ρ€Π°Ρ‚ΠΈΡ‚ΡŒ дСньги ΠΈΠ»ΠΈ врСмя.
  • Команда ΠΈ Ρ€Π°Π±ΠΎΡ‡ΠΈΠΉ процСсс. ΠšΠΎΠΌΠ°Π½Π΄Ρ‹ с ΠΎΠ±Ρ‰ΠΈΠΌ ΠΏΠΎΠ½ΠΈΠΌΠ°Π½ΠΈΠ΅ΠΌ ΠΏΡ€ΠΈΠ½ΠΈΠΌΠ°ΡŽΡ‚ Π±ΠΎΠ»Π΅Π΅ эффСктивныС Ρ€Π΅ΡˆΠ΅Π½ΠΈΡ ΠΏΠΎ ΠΏΡ€ΠΎΠ΄ΡƒΠΊΡ‚Π°ΠΌ, ΠΏΠΎΠ»ΠΈΡ‚ΠΈΠΊΠ΅ ΠΈ ΠΎΠ±ΡƒΡ‡Π΅Π½ΠΈΡŽ.

Π‘Π»Π΅Π΄ΠΈΡ‚Π΅ Π·Π°

  • Π Π°Π·Π½Ρ‹Π΅ ΠΊΠΎΠΌΠ°Π½Π΄Ρ‹ ΠΌΠΎΠ³ΡƒΡ‚ ΠΈΡΠΏΠΎΠ»ΡŒΠ·ΠΎΠ²Π°Ρ‚ΡŒ ΠΎΠ΄ΠΈΠ½ ΠΈ Ρ‚ΠΎΡ‚ ΠΆΠ΅ Ρ‚Π΅Ρ€ΠΌΠΈΠ½ ΠΏΠΎ-Ρ€Π°Π·Π½ΠΎΠΌΡƒ, поэтому Π·Π°Ρ€Π°Π½Π΅Π΅ ΠΎΠΏΡ€Π΅Π΄Π΅Π»ΠΈΡ‚Π΅ ΠΌΠ°ΡΡˆΡ‚Π°Π±.
  • ВСсты ΠΌΠΎΠ³ΡƒΡ‚ Π²Ρ‹Π³Π»ΡΠ΄Π΅Ρ‚ΡŒ ΡΠΈΠ»ΡŒΠ½Ρ‹ΠΌΠΈ, Π² Ρ‚ΠΎ врСмя ΠΊΠ°ΠΊ Ρ€Π΅Π°Π»ΡŒΠ½Π°Ρ ΠΏΡ€ΠΎΠΈΠ·Π²ΠΎΠ΄ΠΈΡ‚Π΅Π»ΡŒΠ½ΠΎΡΡ‚ΡŒ Π½Π΅Ρ€Π°Π²Π½ΠΎΠΌΠ΅Ρ€Π½Π°.
  • Π˜Π³Π½ΠΎΡ€ΠΈΡ€ΠΎΠ²Π°Π½ΠΈΠ΅ качСства Π΄Π°Π½Π½Ρ‹Ρ… ΠΈ ΠΏΠ»Π°Π½ΠΎΠ² ΠΎΡ†Π΅Π½ΠΊΠΈ часто ΠΏΡ€ΠΈΠ²ΠΎΠ΄ΠΈΡ‚ ΠΊ Π½Π΅ΡΡ‚Π°Π±ΠΈΠ»ΡŒΠ½Ρ‹ΠΌ Ρ€Π΅Π·ΡƒΠ»ΡŒΡ‚Π°Ρ‚Π°ΠΌ.

Π‘Π»Π΅Π΄ΡƒΡŽΡ‰ΠΈΠ΅ шаги

  1. НачнитС с простого опрСдСлСния ΠΆΠ΅Π»Π°Π΅ΠΌΠΎΠ³ΠΎ Ρ€Π΅Π·ΡƒΠ»ΡŒΡ‚Π°Ρ‚Π°.
  2. ΠŸΠ΅Ρ€Π΅Π΄ тСстированиСм Π²Ρ‹Π±Π΅Ρ€ΠΈΡ‚Π΅ ΠΎΠ΄ΠΈΠ½ ΠΏΠΎΠΊΠ°Π·Π°Ρ‚Π΅Π»ΡŒ успСха ΠΈ ΠΎΠ΄Π½ΠΎ условиС ΠΎΡ‚ΠΊΠ°Π·Π°.
  3. ЗапуститС нСбольшой ΠΏΠΈΠ»ΠΎΡ‚Π½Ρ‹ΠΉ ΠΏΡ€ΠΎΠ΅ΠΊΡ‚ с Ρ€Π΅ΠΏΡ€Π΅Π·Π΅Π½Ρ‚Π°Ρ‚ΠΈΠ²Π½Ρ‹ΠΌΠΈ Π΄Π°Π½Π½Ρ‹ΠΌΠΈ, Π° Π½Π΅ ΠΎΡ‚Ρ‚ΠΎΡ‡Π΅Π½Π½Ρ‹ΠΉ дСмонстрационный Π½Π°Π±ΠΎΡ€.
  4. Π”ΠΎΠΊΡƒΠΌΠ΅Π½Ρ‚ΠΈΡ€ΡƒΠΉΡ‚Π΅, Π³Π΄Π΅ ΠΏΠΎΠΌΠΎΠ³Π°ΡŽΡ‚ Π³Ρ€ΠΎΠΊΠΊΠΈΠ½Π³ ΠΈ ΠΎΡ‚Π»ΠΎΠΆΠ΅Π½Π½ΠΎΠ΅ ΠΎΠ±ΠΎΠ±Ρ‰Π΅Π½ΠΈΠ΅ ΠΈ Π³Π΄Π΅ Π±ΠΎΠ»Π΅Π΅ простыС ΠΌΠ΅Ρ‚ΠΎΠ΄Ρ‹ Π»ΡƒΡ‡ΡˆΠ΅.

ΠŸΡ€ΠΎΠ΄ΠΎΠ»ΠΆΠ°ΠΉΡ‚Π΅ ΠΈΡΡΠ»Π΅Π΄ΠΎΠ²Π°Ρ‚ΡŒ

БСсплатная рассылка

ΠŸΠΎΠ»ΡƒΡ‡Π°ΠΉΡ‚Π΅ Π΅ΠΆΠ΅Π΄Π½Π΅Π²Π½Ρ‹ΠΉ AI-ΠΎΠ±Π·ΠΎΡ€

Π’Ρ€ΠΈ ΠΏΡ€ΠΎΠ²Π΅Ρ€Π΅Π½Π½Ρ‹Ρ… AI-истории ΠΊΠ°ΠΆΠ΄ΠΎΠ΅ ΡƒΡ‚Ρ€ΠΎ Π±ΡƒΠ΄Π½Π΅Π³ΠΎ дня, написанныС простым английским. БСсплатно навсСгда, Π±Π΅Π· Ρ€Π΅ΠΊΠ»Π°ΠΌΡ‹.

Одно элСктронноС письмо ΠΊΠ°ΠΆΠ΄Ρ‹ΠΉ Π±ΡƒΠ΄Π½ΠΈΠΉ дСнь. ΠžΡ‚ΠΏΠΈΡΠ°Ρ‚ΡŒΡΡ Π² ΠΎΠ΄ΠΈΠ½ ΠΊΠ»ΠΈΠΊ. ΠœΡ‹ Π½ΠΈΠΊΠΎΠ³Π΄Π° Π½Π΅ ΠΏΡ€ΠΎΠ΄Π°Π΅ΠΌ ΠΈ Π½Π΅ ΠΏΠ΅Ρ€Π΅Π΄Π°Π΅ΠΌ ваш адрСс.

ΠŸΡ€ΠΎΠ²Π΅Ρ€ΡŒΡ‚Π΅ сСбя

Take the Grokking and Delayed Generalization quiz

МгновСнная обратная связь ΠΏΠΎ ΠΊΠ°ΠΆΠ΄ΠΎΠΌΡƒ ΠΎΡ‚Π²Π΅Ρ‚Ρƒ ΠΈ ΠΎΠ±Ρ‰ΠΈΠΉ сСртификат с провСряСмым ΠΈΠ΄Π΅Π½Ρ‚ΠΈΡ„ΠΈΠΊΠ°Ρ‚ΠΎΡ€ΠΎΠΌ послС прохоТдСния курса.

ΠΠ°Ρ‡Π°Ρ‚ΡŒ тСст

ΠŸΠΎΠ΄Π΄Π΅Ρ€ΠΆΠΈΡ‚Π΅ бСсплатноС ΠΎΠ±Ρ€Π°Π·ΠΎΠ²Π°Π½ΠΈΠ΅ Π² области искусствСнного ΠΈΠ½Ρ‚Π΅Π»Π»Π΅ΠΊΡ‚Π°. AI Understanding β€” это нСкоммСрчСская организация 501(c)(3). Никакой Ρ€Π΅ΠΊΠ»Π°ΠΌΡ‹ ΠΈ ΠΏΠ»Π°Ρ‚Π½ΠΎΠ³ΠΎ доступа. Π‘Π΄Π΅Π»Π°Ρ‚ΡŒ ΠΏΠΎΠΆΠ΅Ρ€Ρ‚Π²ΠΎΠ²Π°Π½ΠΈΠ΅

Часто Π·Π°Π΄Π°Π²Π°Π΅ΠΌΡ‹Π΅ вопросы

What is Grokking and Delayed Generalization?

Grokking describes studied cases where held-out generalization improves well after a network has fit its training examples. Small algorithmic tasks provide clear experiments, but neither a validation plateau nor longer training guarantees that this transition will occur.

Which performance pattern characterizes the studied grokking phenomenon?

The delay concerns generalization beyond observed examples under the studied setup.

Which tasks did the original grokking paper use as a controlled setting?

The public paper was submitted in January 2022 and studies small algorithmic datasets.

Which regularization intervention was among those examined in the original grokking experiments?

The original study investigates weight decay and other optimization choices; no single universal cause is established.

What computation did the cited modular-addition mechanistic study identify in its trained small transformers?

Analysis and Fourier ablations support the particular studied algorithm.

Which phase ordering does Progress Measures for Grokking describe in the studied training dynamics?

The generalizing circuit forms before the later cleanup associated with the visible test-accuracy transition.

ΠŸΡ€ΠΎΠ΄ΠΎΠ»ΠΆΠ°ΠΉΡ‚Π΅ ΡƒΡ‡ΠΈΡ‚ΡŒΡΡ

БвязанныС руководства

Π”Ρ€ΡƒΠ³ΠΈΠ΅ руководства, Π²Ρ‹Π±Ρ€Π°Π½Π½Ρ‹Π΅ ΠΏΠΎ этой Ρ‚Π΅ΠΌΠ΅

ΠŸΠΎΠ»ΡƒΡ‡ΠΈΡ‚Π΅ бСсплатный AI-ΡΠ΅Ρ€Ρ‚ΠΈΡ„ΠΈΠΊΠ°Ρ‚ΠšΠΎΡ€ΠΎΡ‚ΠΊΠΈΠΉ бСсплатный курс ΠΈ экзамСн. Π‘Π΅Π· стоимости, Π±Π΅Π· ΠΊΠ°Ρ€Ρ‚Ρ‹. Около 2,5 часов.ΠΠ°Ρ‡Π°Ρ‚ΡŒ бСсплатно β†’