Skip to content

Minimal recurrent motif (sb1 rs2 g0.18) – non-record submission#323

Open
megnat05-tmm wants to merge 5 commits intoopenai:mainfrom
megnat05-tmm:main
Open

Minimal recurrent motif (sb1 rs2 g0.18) – non-record submission#323
megnat05-tmm wants to merge 5 commits intoopenai:mainfrom
megnat05-tmm:main

Conversation

@megnat05-tmm
Copy link

This is a non-record submission.

Summary

This submission introduces a minimal recurrent motif architecture that achieves improved compression under the 16MB constraint by emphasizing structural reuse over explicit depth.

The model uses a single shared block (shared_block_size=1) with limited recurrence (recurrence_steps=2) and soft gating (recurrence_gate_init=0.18). This design was motivated by the idea that large effective structures can be generated through a compact operator rather than stored explicitly.

Results

Final (roundtrip):

  • val_loss: 4.7062
  • val_bpb: 2.7873

Artifact:

  • compressed size: ~1.92 MB
  • raw size: ~8.47 MB

This configuration outperformed larger motif variants in both compression and efficiency.

Approach

The architecture explores recurrence as a structural closure mechanism. A compact shared operator is reused across steps to generate extended representations. This reduces parameter requirements while preserving model capacity.

Notes

  • Validation timing on local hardware reflects evaluation chunking and logging cadence, but does not affect correctness of metrics.
  • The submission is fully self-contained and reproducible.

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