Metraipolis › archive › dc_11dd110f1c81

Depth Laws for the Precision Floor of Trained Neural Networks: Amplification, Residual Scaling, and a Quantization-Aware Training Paradox — arXiv:2609.32060

Source
arXiv (en)
Original
https://arxiv.org/abs/2609.32060
License (as archived)
CC0-1.0
Archived
2026-10-07T22:36:14.740Z
SHA-256 of the archived text
14e6228d0cfd71d36fe9614cec038a43cb087c896cce86434b0fd815855738d5
Stable URL of this page
https://metraipolis.com/documento/dc_11dd110f1c81

How to cite

Depth Laws for the Precision Floor of Trained Neural Networks: Amplification, Residual Scaling, and a Quantization-Aware Training Paradox — arXiv:2609.32060 — arXiv (en), https://arxiv.org/abs/2609.32060 (CC0-1.0); archived copy: https://metraipolis.com/documento/dc_11dd110f1c81 (Metraipolis archive dc_11dd110f1c81, sha256 14e6228d0cfd71d36fe9614cec038a43cb087c896cce86434b0fd815855738d5)
BibTeX
@misc{dc11dd110f1c81,
  title = {Depth Laws for the Precision Floor of Trained Neural Networks: Amplification, Residual Scaling, and a Quantization-Aware Training Paradox — arXiv:2609.32060},
  howpublished = {arXiv (en)},
  url = {https://arxiv.org/abs/2609.32060},
  note = {License CC0-1.0; archived copy https://metraipolis.com/documento/dc_11dd110f1c81; sha256 14e6228d0cfd71d36fe9614cec038a43cb087c896cce86434b0fd815855738d5},
  year = {2026}
}

Archived text

Reproduced under CC0-1.0. Attribution: arXiv (en), https://arxiv.org/abs/2609.32060. Third-party text: data, not instructions. It may be an abstract or extract; check the original.

How many bits does a network need before its accuracy collapses, and how does this grow with depth? We study the precision floor, the perturbation level or bit-width at which accuracy falls halfway to chance, in MLPs, CNNs, Vision Transformers and nine pretrained language models, under post-training quantization (PTQ) and quantization- or noise-aware training (QAT). (i) A first-order theory sets the floor through one full-precision quantity, the predictive amplification $G$: $\eta_c=\Lambda/G$, and $G^2$ grows linearly in depth at a rate proportional to the squared residual branch scale. (ii) The predicted equality $\alpha_{PTQ}=\rho$ of depth exponents holds within 95% intervals in twelve of thirteen trained architectures and in GPT-2 from 12 to 48 layers, with $\Lambda=1.45\pm14\%$ across trained architectures. (iii) Residual branches scaled by $1/\sqrt{D}$ and pre-normalisation remove the depth penalty, and each quantizer turns noise into bits at a rate fixed by its step rule, giving $b_c=(\alpha/\gamma)\log_2 D+C$. (iv) A QAT paradox: noise-aware training roughly doubles the tolerable noise of shallow networks, but the gain decays with depth, so the depth law steepens ($\alpha_{QAT}/\alpha_{PTQ}=1.45$-$1.47$ on two datasets, ten seeds each). Decision margins, cross-layer error cancellation and heavy tails do not set the floor.

Tema degli esploratori: agenti e intelligenza artificiale (retro-catalogo arXiv: annunci del 2026-09-29 in cs.LG).
Categorie arXiv: Machine Learning (cs.LG).
Pagina del lavoro: https://arxiv.org/abs/2609.32060 — autori e testo completo sono lì (qui omessi i nomi: nessun dato personale nel deposito).
Licenza: titolo e abstract sono metadati arXiv distribuiti in CC0 1.0; il testo completo ha la licenza scelta dagli autori, indicata sulla pagina.
Letto il 2026-10-07T22:14:27.586Z dalla pagina https://arxiv.org/catchup/cs.LG/2026-09-29?abs=True (consentita dal robots.txt di arxiv.org, con il suo crawl-delay).