Conditioned Direct Feedback Alignment

Code and evidence for Conditioned Direct Feedback Alignment via Activity and Error Geometry, by Houman Safaai, Varun Reddy, and Bernardo L. Sabatini.

The September 26 revision includes the completed joint confirmation: 880 training runs across eight cohorts, ten paired seeds, and 21 primary contrasts. Activity conditioning and its established FOOF formulation improve tuned DFA under matched measured training time. Covariance controls and early-only conditioning support more specific conclusions about activity statistics and training cost. The stable-training error-factor gate fails despite favorable mean effects; the timing tests do not establish a DFA-specific mechanism, and the image-background predictions are not confirmed. Equal measured time does not imply equal FLOPs.

The paper repository contains the current manuscript, figures, and checked ICLR/arXiv build commands. The arXiv record is updated separately; preparing a source bundle does not replace that public record.

The focused arXiv correction release is tagged ndfa-arxiv-review-2026-09-26-v2 in both repositories. Its revision guide maps the frozen protocol, recorded timestamps and hashes, current figure inputs, and publicly available evidence. It clarifies conditioning before AdamW and corrects presentation without changing the results. The second correction pass distinguishes the separate noninferiority criterion, narrows the Fisher summary, and repairs notation and prose interruptions; figure data are unchanged.

Verify the current results

python -m pip install -r requirements.txt
python scripts/ndfa_strengthening_20260925/audit_joint_decision.py \
  --root assets/ndfa_confirmation_20260926 \
  --output build/confirmation_verification.json

This verifies the exported file hashes, all 880 declared cases, all 21 primary seed-level contrasts, individual t intervals, global Holm adjustment, and the error-factor and efficiency gates. It uses supplied endpoint summaries and does not retrain models or reevaluate checkpoints. See REPRODUCE.md for CPU checks, source builds, and relocation of the frozen training protocol.

Evidence or implementation	Location
Frozen settings, test summaries, source snapshot, and claim map	Confirmation evidence
Activity and error conditioning	Local preconditioning, integrated operators
FOOF with either credit rule	FOOF, training loop
Forward-decorrelation reference	FD implementation
Validation selection, test gate, and declared comparisons	Study control
Independent endpoint reconstruction	Audit
Earlier corrected theory and cohorts	Mathematical checks, September 25 evidence

The public evidence contains every retained confirmation endpoint, including direction diagnostics. Dataset files, model checkpoints, and per-example prediction tensors are not included. Historical absolute paths and hashes in the frozen records identify the original experiment; the portable audit and rerun helper use local copies. Historical development notes describe their own stages and must not replace the completed confirmation's conclusions.

Citation

@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}
}