Code for Conditioned Direct Feedback Alignment via Activity and Error Geometry, by Houman Safaai, Varun Reddy, and Bernardo L. Sabatini. The manuscript repository contains the anonymous ICLR entrypoint and author preprint sources. The published arXiv preprint predates the September 18–19 scientific revision described here.
Activity conditioning improves fixed-feedback learning in nuisance-dominated settings and retains gains over DFA with shared normalization, wider MLPs and matched measured work. Error conditioning provides a distinct benefit in the clean factor confirmations but depends more strongly on damping and protocol. BP remains stronger in the matched-work comparisons; forward decorrelation is competitive. The paper studies update geometry under approximate credit, with explicit limits on biological locality and practical generalization.
The current second-review revision qualifies the full-versus-diagonal moment interpretation, fixes the remaining seed-level uncertainty summaries, and clarifies the theory's joint limit and flow clock. It organizes the supplement into 13 appendices, removes duplicate/development presentations and supplies a study ledger. The active paper has 14 figures and 49 references. No new model training was needed. See the second-review findings.
The current figure revision compacts Figure 3 and expands Figure 5 to four panels, adding the existing work and width follow-ups. Duplicate supplementary panels are removed. The preceding horizontal layouts in Figures 1, 2 and 4 remain unchanged; corrected measurements and uncertainty estimates are preserved.
The preceding experimental revision remains integrated, including all 202 targeted follow-up cases. Original experimental records remain available in the research and review archives.
REPRODUCE.md gives build, verification, figure and training
commands. The artifact map
binds source hashes, audited evidence, PDFs and upload archives to this revision.
The current tag is ndfa-expanded-figures-2026-09-19 in both code and manuscript repositories.
| Claim or artifact | Implementation / record |
|---|---|
| Complete-risk calculation and seed-level resampling | Mathematical functions, mode-timing simulation, seed analysis |
| Longer matched work, width-specific damping and BN intervention | Frozen protocol, configuration, runner |
| Saved-logit endpoints, loss tails and existing diagonal control | New-study audit, saved-evidence audit |
| Earlier matched-work and forward-decorrelation comparison | Protocol, training and verification |
| Controlled moment orientation and credit quality | Experiment, analysis |
| Seed-level BN / CIFAR-100 corrections and cohort provenance | Checks, study ledger |
| Revised publication figures | Compact overlays, base generator, corrected numerical checks |
The compact anonymous evidence archive is prepared as a separate submission
attachment. Generated PDFs, the archive, datasets, raw logits and checkpoints
are not stored in this Git repository. Paths under results/ in the artifact
map identify the preserved workspace/review-package evidence, not public GitHub
files. The compact archive verifies reported endpoints from saved per-example
losses and classes; it does not rerun GPU training, checkpoint inference or timing.
No arXiv replacement or conference submission was performed during this revision. Earlier development documentation is historical; use the revision findings and artifact map for current claims.
@misc{safaai2026conditioned,
title = {Conditioned Direct Feedback Alignment via Activity and Error Geometry},
author = {Safaai, Houman and Reddy, Varun and Sabatini, Bernardo L.},
year = {2026},
eprint = {2607.18574},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2607.18574}
}