diff options
| -rw-r--r-- | index.html | 147 | ||||
| -rw-r--r-- | robots.txt | 2 |
2 files changed, 149 insertions, 0 deletions
diff --git a/index.html b/index.html new file mode 100644 index 0000000..eaa8f7c --- /dev/null +++ b/index.html @@ -0,0 +1,147 @@ +<!DOCTYPE html> +<html lang="en"> +<head> +<meta charset="utf-8"> +<meta name="viewport" content="width=device-width, initial-scale=1"> +<meta name="robots" content="noindex, nofollow, noarchive"> +<title>Training Transformer Language Models Without Backpropagation</title> +<style> + :root { --ink:#1a1a1a; --muted:#555; --accent:#8a1f1f; --rule:#e5e0d8; --bg:#faf8f4; } + html { background: var(--bg); } + body { max-width: 780px; margin: 0 auto; padding: 48px 20px 80px; + font: 17px/1.65 Georgia, 'Times New Roman', serif; color: var(--ink); } + h1 { font-size: 34px; line-height: 1.25; margin: 8px 0 10px; letter-spacing: -0.3px; } + .sub { font-size: 19px; color: var(--muted); margin-bottom: 18px; } + .authors { color: var(--muted); margin: 6px 0 4px; } + .links { margin: 18px 0 8px; } + .links a, .links span { display: inline-block; border: 1px solid var(--ink); border-radius: 20px; + padding: 4px 16px; margin: 0 6px 8px 0; font-family: Helvetica, Arial, sans-serif; + font-size: 14px; text-decoration: none; color: var(--ink); } + .links span { color: #999; border-color: #ccc; } + h2 { font-size: 22px; margin: 40px 0 10px; border-bottom: 2px solid var(--rule); padding-bottom: 6px; } + .tldr li { margin: 9px 0; } + b.k { color: var(--accent); } + table { border-collapse: collapse; width: 100%; margin: 14px 0; font-size: 15.5px; } + th, td { text-align: left; padding: 7px 10px; border-bottom: 1px solid var(--rule); } + th { font-family: Helvetica, Arial, sans-serif; font-size: 13px; text-transform: uppercase; + letter-spacing: 0.04em; color: var(--muted); } + .note { font-size: 14.5px; color: var(--muted); } + .rule-claim { background: #f3efe7; border-left: 4px solid var(--accent); padding: 12px 16px; margin: 18px 0; } + code, pre { font: 13.5px/1.5 ui-monospace, 'SF Mono', Menlo, monospace; } + pre { background: #f3efe7; padding: 14px 16px; overflow-x: auto; border-radius: 6px; } + footer { margin-top: 60px; font-size: 13.5px; color: var(--muted); border-top: 1px solid var(--rule); padding-top: 14px; } +</style> +</head> +<body> + +<h1>Training Transformer Language Models Without Backpropagation</h1> +<div class="sub">Standard LLMs, trained end-to-end by Equilibrium Propagation — with matched backprop controls at every scale, and every design choice made under analog-hardware constraints.</div> + +<div class="authors">Author list forthcoming · University of Illinois Urbana-Champaign</div> +<div class="links"> + <span>Paper — in preparation</span> + <span>Code — on request</span> + <span>Training curves — on request</span> +</div> + +<h2>TL;DR</h2> +<ul class="tldr"> + <li><b class="k">The largest models ever trained from scratch without backpropagation at any level.</b> + Every parameter update follows the Equilibrium Propagation (EP) rule; no layer, block, or output head + is trained with a backprop rule. Our models are standard transformer LMs (OLMo2-style blocks, + 32k-token vocabulary, FineWeb-Edu) — the trained network runs ordinary forward inference.</li> + <li><b class="k">Matches backprop within a 4–5% perplexity band at 72M parameters</b>, against a + backprop twin trained on identical data, steps, and optimizer — with multi-seed controls on both sides.</li> + <li><b class="k">First transformer language model in the EP family, and 2.15× the size of the + largest prior EP-family result</b> (VGG10 image classifier, ∼63M params, Kerjan–Høier–Scellier 2026): + models trained up to <b>135M</b> parameters on 2.7B tokens.</li> + <li><b class="k">2× wall-clock overhead versus backprop — where the closest prior EP work pays + 6.7–12×.</b> One nudged phase and 3 relaxation sweeps per step versus their two phases × 10 + iterations: 6.7× fewer relaxation iterations per step.</li> + <li><b class="k">Verified gradient fidelity, not alignment by luck:</b> the EP update maintains cosine + ≈0.99 to the true backprop gradient throughout training — measured, at scale. Prior non-backprop + transformer results at larger sizes train most parameters with <i>local backprop inside blocks</i>; + ours use none.</li> + <li><b class="k">Hardware-ready by measurement, not assumption:</b> 8-bit weight/compute quantization shows + <i>no EP-specific penalty</i> against an equally-quantized backprop twin; under injected analog faults + (1% forward noise, 10% error-channel noise, device-tolerance mismatch) the EP estimator tracks the + faulted network at cosine ≈0.97 — learning co-adapts to the hardware.</li> +</ul> + +<h2>Why this matters</h2> +<p>Analog and physical accelerators promise order-of-magnitude energy savings for training, but they cannot +run backpropagation natively: exact gradients on a physical substrate require either per-step digitization +or a second, adjoint copy of the hardware. Equilibrium Propagation extracts gradients from the physics +itself — two relaxations and local reads — and is the only member of its family with a +gradient-equivalence guarantee we can verify at scale. The missing evidence has always been scale and +rigor: EP results stopped at mid-size vision models, without matched controls. This project supplies both, +on the model class that matters commercially: language models.</p> + +<div class="rule-claim"> +<b>The claim we defend:</b> a standard transformer LM can be trained to backprop-class quality with a +physics-compatible learning rule, at 2× backprop wall-clock in simulation — and the constraints that +matter for analog hardware (quantization, noise, nudge operating windows, energy) are measured quantities +in our stack, not assumptions. +</div> + +<h2>Headline results</h2> +<table> + <tr><th>Model</th><th>Data</th><th>EP (val CE)</th><th>BP twin</th><th>Gap</th></tr> + <tr><td>72M transformer LM</td><td>FineWeb-Edu, 1.4B tokens</td><td>3.33</td><td>3.29</td> + <td>+4–5% ppl</td></tr> + <tr><td>135M transformer LM</td><td>FineWeb-Edu, 2.7B tokens</td><td colspan="3">trained end-to-end, zero + instability events; scaling analysis below</td></tr> +</table> +<p class="note">Twin discipline: identical architecture, tokenizer, data order, optimizer (Muon hybrid), +steps, and evaluation; multi-seed on both sides (BP n=3, band ±0.006; EP n=2). Comparison row uses the +sealed flagship pair.</p> + +<table> + <tr><th>Cost vs backprop (wall-clock)</th><th>This work</th><th>Closest EP work (ImageNet VGG10)</th></tr> + <tr><td>Training overhead</td><td><b>2.0×</b></td><td>6.7× (single-sided) / 12× (centered)</td></tr> + <tr><td>Relaxation iterations / step</td><td><b>3</b> (one phase)</td><td>20 (two phases × 10)</td></tr> +</table> + +<h2>The scaling science</h2> +<p>Scaling a physical learning rule surfaces phenomena backprop never meets. Between 72M (width 512) and +135M (width 768) we identified a width-scaling loss in the EP gradient — localized to the top half of +the network, invisible to every per-step alignment metric, and traced to response components that finite +nudge displacement under-reaches. We built a screening instrument that measures this leak in 90 minutes +per candidate recipe, mapped its dose–response law, and demonstrated a pure estimator-side treatment +that recovers 97% of it without touching the model, its inference path, or the cost budget. Full-schedule +validation of the treated recipe is running now; the same instruments give the go/no-go protocol for each +next rung of the ladder (300M → 1B).</p> + +<h2>Hardware line</h2> +<ul> + <li>Measured energy projection for an integrated weight-stationary realization: <b>0.21–0.63 pJ/MAC</b> + (SPICE-measured analog core + datasheet periphery), against a 0.3–1 pJ/MAC digital INT8 system envelope.</li> + <li>Single-column analog prototype: SPICE-modeled, discrete multiplying-DAC parts list — deliberately + kept at the “hardware someone can actually build” level.</li> + <li>Nudge-amplitude operating windows and their evolution over training are mapped — the dynamic-range + spec an analog implementation must meet.</li> +</ul> + +<h2>Roadmap</h2> +<p>Staged scaling with matched BP controls and hardware-relevant ablations at every rung: 150M–600M +scaling ladder (does the gap grow or shrink with scale — measured, not assumed), then 1B–3B. +Each stage is gated on the previous stage’s loss, alignment, and throughput numbers. In parallel: the +algorithm→regime map across the activity-difference family (contrastive / coupled-learning arms on the +same harness), and the single-column hardware feasibility study.</p> + +<h2>BibTeX</h2> +<pre>@misc{ept2026, + title = {Training Transformer Language Models Without Backpropagation}, + author = {(author list forthcoming)}, + year = {2026}, + note = {Project page} +}</pre> + +<footer> +Contact: yurenh2@illinois.edu · This page is intentionally excluded from search indexing while the +work is under review. Numbers on this page correspond to sealed experiment records with matched controls; +the technical report contains the full protocol. +</footer> + +</body> +</html> diff --git a/robots.txt b/robots.txt new file mode 100644 index 0000000..1f53798 --- /dev/null +++ b/robots.txt @@ -0,0 +1,2 @@ +User-agent: * +Disallow: / |