Skip to content

[FEATURE] neuron Phase 16 — dynamic topology evolution (RigL regrow + Net2Net grow) [메인] #75

Description

@juhy0987

배경 (Phase 15 직후)

Phase 15 (PR #72 머지) 에서 edge prune (dynamic shrink) 도입 — paradigm 의 진짜 dynamic phase 첫 단계. 검증 결과:

  • 30% prune almost-free (+0.041 loss), 70% 도 catastrophic cliff 없음 (+0.134)
  • post-prune recovery curve 가 gradient resurrection 방지 메커니즘의 정확성을 학습 dynamics 차원에서 입증

Phase 15 의 graceful capacity scaling 가능성 → Phase 16 의 두 갈래 시도:

Sub-Phase 메커니즘 핵심 가설
16a (RigL/SET DST) iterative prune + regrow, constant sparsity 유지 30% almost-free 구간이 topology 재할당 으로 50%+ 까지 확장 가능?
16b (Net2Net grow) pruned slot 에 function-preserving 신규 edge 추가 (또는 shape 확장) 학습 중 capacity 동적 확장 가능? Phase 15 prune 의 반대 방향.

두 갈래 모두 paradigm 의 dynamic phase 의 다른 차원:

  • 16a: within-shape dynamics (sparsity reallocation, total 일정)
  • 16b: cross-shape dynamics (capacity expansion, total 변동)

Sub-issue 분할 (sub-issue Relation 활성화 예정)

각 sub-issue 가 별도 PR 로 진행. 본 메인 이슈는 두 sub-issue 모두 close 될 때 함께 close.

완료 조건

  • Sub-issue 16a 머지 (RigL regrow + 노트북 + verdict)
  • Sub-issue 16b 머지 (Net2Net grow + 노트북 + verdict)
  • Notion Phase 16 종합 페이지 작성 (두 sub-phase 결과 통합)

관련

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    enhancementNew feature or request

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions