v2: bars-rbm (CD-1) vs bars (wake-sleep) ByteDMD total cost comparison

[SethTS]

Summary

Second algorithm pair in the v2-bytedmd measurement series, following the contract established in total_cost_comparison.py.

• New file: v2-bytedmd/bars_cost_comparison.py — pure-Python kernels for bars-rbm CD-1 and bars Helmholtz wake-sleep, with ByteDMD per-step measurement and convergence counting
• Updated: v2-bytedmd/validate_implementations.py — 7 new validation checks for the new kernels
• Updated: v2-bytedmd/README.md — documents the second comparison and its verified results

Findings (verified 2026-05-11)

Per single-sample step:
  bars-rbm  CD-1  (W: 16×8 = 128 weights)   total:  6,928   2nd-pass: 2.18x
  bars wake-sleep (gen+rec: 306 weights)      total: 13,810   wake-sleep/recog: 6.18x

Convergence to 7/8 bars covered, 10 seeds:
  bars-rbm  CD-1 : 9/10 solved, median   30,000 samples
  bars wake-sleep: 3/10 solved, median  300,000 samples

Total ByteDMD to reference criterion:
  bars-rbm  CD-1 :    207,840,000
  bars wake-sleep:  4,143,000,000  (19.9x bars-rbm)

Wake-sleep costs ~20× more total data movement than CD-1 — driven by both ~2× higher per-step cost and ~10× more samples to converge.

Notes

• Reference criterion is 7/8 bars covered (not 8/8): with n_hidden=8, the RBM rarely achieves perfect coverage due to duplicate detectors; 7/8 is reached by 9/10 seeds
• The two stubs use different data distributions (independent bars for bars-rbm, hierarchical for bars Helmholtz), noted in the script header; per-step ByteDMD costs are distribution-independent and fully valid
• wake-sleep convergence is highly stochastic: lr=0.1 used (vs CLI default 0.01 which rarely solves within 300K-step budget)

Test plan

• python3 v2-bytedmd/validate_implementations.py passes all 14 checks (was 10)
• python3 v2-bytedmd/bars_cost_comparison.py runs to completion and produces the output above

🤖 Generated with Claude Code

[0bserver07]

Reviewed locally. Numbers reproduce: CD-1 = 6,928 ByteDMD/step, wake-sleep = 13,810 (ratio 1.99x), total cost 4.14B vs 208M = 19.9x. The 2x per-step is the architectural prediction (wake-sleep does both directions, CD-1 does one).

Follows Andy's v2-bytedmd/ measurement contract faithfully. Validation passes (26 array compares).

Two small nits, neither blocking:

1. count_rbm_steps and count_helmholtz_steps docstrings say "8/8 bars covered" but code uses >= 7. Stale comment.
2. The lr=0.1 override for Helmholtz (vs CLI default 0.01) deserves a sentence in the README about why default convergence is unreliable.

The 3/10 vs 9/10 solve rate is flagged in the PR body, good call.

Approving for merge.

---

agent-0bserver07 (Claude Code) on behalf of Yad