Skip to content

feat(ocl): add Synaptic Intelligence (SI) strategy - #403

Open
tachyonicClock wants to merge 1 commit into
mainfrom
feat/ocl-si
Open

feat(ocl): add Synaptic Intelligence (SI) strategy#403
tachyonicClock wants to merge 1 commit into
mainfrom
feat/ocl-si

Conversation

@tachyonicClock

@tachyonicClock tachyonicClock commented Aug 24, 2026

Copy link
Copy Markdown
Collaborator
  • Cherry-picks the Synaptic Intelligence (SI) continual learning strategy from the blurry-ocl draft branch, adapting it to the current capymoa.ocl module layout (capymoa.ocl.events, capymoa.ocl.util._buffer_list).
  • Adds reset_optimizer_state, a small utility SI uses to clear optimiser momentum at task boundaries.
  • Registers SI in capymoa.ocl.strategy.__init__.
  • Adds masked and unmasked SI cases to tests/ocl/test_strategy.py, extending new_constructor with an optional optimiser_type (defaults to SGD, backwards compatible) since SI's reference hyperparameters use Adam.

Closes: Synaptic Intelligence

@tachyonicClock tachyonicClock changed the title feat/ocl si feat(ocl): add Synaptic Intelligence (SI) strategy Aug 24, 2026
@tachyonicClock
tachyonicClock marked this pull request as ready for review August 24, 2026 00:47
Synaptic Intelligence (SI) is a regularisation-based continual learning
strategy that accumulates per-parameter importance online from
optimisation trajectories, then penalises changes to parameters that
were important for previous tasks (Zenke et al., 2017).

Ported from the `blurry-ocl` draft branch and adapted to the current
`capymoa.ocl` module layout. Also fixes the `update_trajectory`
parameter names, which had drifted from their actual pre/post-step
meaning during earlier draft cleanup, and adds `reset_optimizer_state`,
a small utility SI needs to clear optimiser momentum at task
boundaries.

Assisted-by: claude-code:claude-sonnet-5
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant