v2: ByteDMD / data-movement instrumentation on v1+v1.5 baselines

[0bserver07]

v2 — ByteDMD / data-movement instrumentation on v1+v1.5 baselines

Why

v1+v1.5 (PRs #4–#16) shipped 58 reproducible synthetic-learning baselines per the spec in #1. The actual research goal was always to measure these under a data-movement / energy lens — that's ByteDMD (reuse-distance proxy for memory-hierarchy energy cost).

The hypothesis (per Yaroslav, hinton-problems issue #1): backprop has a bad commute-to-compute ratio because every gradient step refetches all activations. Algorithmic alternatives that reuse weights more (NBB local rules, fast-weights/linear-attention, capsule routing, etc.) should look better under ByteDMD even when slower in wallclock. v1+v1.5 is the comparator floor; v2 makes the comparison.

This issue mirrors hinton-problems #45, adapted for Schmidhuber's algorithmic lineage.

What this issue tracks

Per-stub instrumentation of the 58 v1+v1.5 implementations with ByteDMD, producing a new column in RESULTS.md (or a dedicated RESULTS_V2.md):

Stub	Run wallclock (v1)	ByteDMD cost (new)	Reuse-distance distribution (new)

Scope

In scope:

• Wrap each stub's train() and final eval() with the ByteDMD tracer
• Output per-stub: total ByteDMD cost, reuse-distance distribution, breakdown by phase (forward / backward / weight update)
• Aggregate into a top-level table sorted by cost, comparable across algorithms

Out of scope:

• Re-implementing any stub's algorithm
• Optimizing for ByteDMD (that's v3 — "find a solver that beats the v1 baseline on data movement")
• Closing the partial reproductions (that's v1.5 follow-up — a sibling issue to this one)

Priority order (per RESULTS.md "v2 filter recommendation")

Tier 1: clean reproductions + sub-second runs (lowest noise floor)

These give the cleanest data-movement signal:

• linear-transformers-fwp (0.08s) — the cleanest pair: equivalent to wave-4's fast-weights-key-value (1992 ancestor); same numpy expression, different schedules. ByteDMD on both should expose whether the equivalence holds at the data-movement level too.
• predictable-stereo (0.08s)
• levin-add-positions (0.34s)
• lococode-ica (0.4s)
• compete-to-compute (0.8s)
• nbb-xor (0.85s)
• rs-two-sequence (0.94s)
• levin-count-inputs (1.0s)
• semilinear-pm-image-patches (1.2s)
• pipe-symbolic-regression (1.3s)
• em-segmentation-isbi (1.5s)
• ssa-bias-transfer-mazes (1.7s)
• chunker-22-symbol (1.86s)
• predictability-min-binary-factors (2.8s)

Tier 2: algorithmic-variant pairs on the same task

These let you compare data-movement properties of different algorithms on identical problems:

• adding-problem family: vanilla RNN vs LSTM (paper's contrast, both implemented in adding-problem and temporal-order-3bit). Direct backprop-cost comparison.
• embedded-reber family: original 1997 LSTM (no forget) vs forget-gate LSTM (continual-embedded-reber). Forget-gate adds compute; does it add commute too?
• LSTM ablation matrix (lstm-search-space-odyssey): 8 variants (V/NIG/NFG/NOG/NIAF/NOAF/CIFG/NP) on the same task. Direct architectural-variant data-movement comparison built in.
• Linear-attention ↔ FWP (linear-transformers-fwp ↔ fast-weights-key-value): the equivalence demo + the 1992 ancestor. ByteDMD on both should produce identical numbers (claim).
• Evolutionary methods: pipe-symbolic-regression (PIPE), evolino-sines-mackey-glass (Evolino), double-pole-no-velocity (ESP), torcs-vision-evolution (DCT-compressed natural ES). Gradient-free family vs gradient-based.
• Search methods: levin-count-inputs, levin-add-positions (Levin), oops-towers-of-hanoi (OOPS), rs-* (random search). All gradient-free; comparable data-movement profile.
• World models: world-models-carracing and world-models-vizdoom-dream share V+M+C decomposition — three distinct training stages with very different memory access patterns.

Tier 3: deferred for v2

Stubs with run wallclock > 100s where v2 ByteDMD overhead would dominate; or partial reproductions where measuring data-movement on a non-converged solver isn't informative:

• pipe-6-bit-parity (240s 6-bit cap)
• evolino-sines-mackey-glass (140s)
• lstm-search-space-odyssey (145s; ablation matrix though, so still high-priority for tier-2 contrast)
• hq-learning-pomdp (paper's HQ-vs-flat gap doesn't reproduce on this maze size; resolve at paper-config first via v1.5)

Acceptance

• ByteDMD cost in RESULTS.md for every stub in tier 1 + tier 2
• A summary plot: ByteDMD cost vs algorithm class (gradient / search / evolutionary / local-rule / fast-weights), x-axis log-scale
• One write-up per algorithmic-variant pair: "X commutes Y× more than Z on the same task; Δ comes from [phase]"

Reference

• ByteDMD: https://github.com/cybertronai/ByteDMD
• Hinton companion: hinton-problems #45

---

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