Context
Tracks three deferred items from the wave 0 implementation of nbb-xor (PR). All three are paper-completeness work, not v1 blockers — the qualitative claim (NBB solves XOR with a strictly local rule) is reproduced in v1.
Three items
1. η / λ rate ablation to close the 5× slowdown
Symptom: paper reports ~619 pattern presentations to converge; we measure ~3012 (mean of 19/20 solvers, single-seed: 3164). Same architecture, same algorithm — the gap is in the rate parameters.
Hypothesis: with η = λ = 0.005, net flow at hidden→correct-output is zero (only redistribution propagates substance). Paper may have used η > λ. Or "presentations" counts differently in the paper.
Ablation: sweep η ∈ {0.005, 0.01, 0.02, 0.05}, λ ∈ {0.005, 0.01}, measure mean presentations to convergence over 30 seeds.
2. 2-hidden + "two 2-unit subsets" architectures
The paper specifies three architectures: 3-hidden (one WTA subset), 2-hidden (impossible by capacity argument), and "two 2-unit subsets". v1 implements only 3-hidden. CLI exposes --n-hidden 2 but the multi-subset case needs a small refactor (per-subset WTA enforcement, separate bucket flows).
3. Source verification beyond IDSIA HTML
v1 reconstructed the rule from IDSIA's HTML transcription (bucketbrigade/node{3,5,6}.html) because the FKI-124-90 PDF on idsia.ch is image-based with degraded OCR. The §Open questions section flags this as a citation gap. v1.5 should:
- Try to retrieve the 1989 Connection Science article in readable form (institutional access)
- Cross-check the §6 algorithm definition against Schmidhuber's 2015 Deep Learning in Neural Networks survey (Section 5.5)
- Flag any divergences from the IDSIA HTML version
Acceptance for this issue
- η/λ sweep run, results posted with chart; if the slowdown is closed, update v1 §Deviations and §Results
- 2-hidden multi-subset implemented and tested; if it works, add to README catalog
- Source check completed; any divergences from IDSIA HTML noted in v1's §Sources
Out of scope
- Energy / data-movement metrics (v2 with ByteDMD)
- Re-implementing other bucket-brigade variants (Sutton credit assignment, etc.)
agent-0bserver07 (Claude Code) on behalf of Yad
Context
Tracks three deferred items from the wave 0 implementation of
nbb-xor(PR). All three are paper-completeness work, not v1 blockers — the qualitative claim (NBB solves XOR with a strictly local rule) is reproduced in v1.Three items
1. η / λ rate ablation to close the 5× slowdown
Symptom: paper reports ~619 pattern presentations to converge; we measure ~3012 (mean of 19/20 solvers, single-seed: 3164). Same architecture, same algorithm — the gap is in the rate parameters.
Hypothesis: with η = λ = 0.005, net flow at hidden→correct-output is zero (only redistribution propagates substance). Paper may have used η > λ. Or "presentations" counts differently in the paper.
Ablation: sweep η ∈ {0.005, 0.01, 0.02, 0.05}, λ ∈ {0.005, 0.01}, measure mean presentations to convergence over 30 seeds.
2. 2-hidden + "two 2-unit subsets" architectures
The paper specifies three architectures: 3-hidden (one WTA subset), 2-hidden (impossible by capacity argument), and "two 2-unit subsets". v1 implements only 3-hidden. CLI exposes
--n-hidden 2but the multi-subset case needs a small refactor (per-subset WTA enforcement, separate bucket flows).3. Source verification beyond IDSIA HTML
v1 reconstructed the rule from IDSIA's HTML transcription (
bucketbrigade/node{3,5,6}.html) because the FKI-124-90 PDF on idsia.ch is image-based with degraded OCR. The §Open questions section flags this as a citation gap. v1.5 should:Acceptance for this issue
Out of scope
agent-0bserver07 (Claude Code) on behalf of Yad