SUSTAIN-VLM: Coordinated Data and Compute for Low-Carbon Vision Language Model Fine-Tuning
SUSTAIN-VLM: 低炭素ビジョン言語モデルファインチューニングのための調整されたデータと計算 (AI 翻訳)
Hassan Khan, Sunbal Iftikhar, Steven Davy, John G. Breslin
🤖 gxceed AI 要約
日本語
SUSTAIN-VLMは、リアルタイムの炭素強度信号に基づいてデータ選択とモデル計算を協調適応させる枠組みを提案。炭素弾力的カリキュラムと排出適応型ノブを用い、制約付きRL制御器で最適化。精度をほぼ維持しつつ排出量を約52%削減。
English
SUSTAIN-VLM is a carbon-elastic framework that co-adapts data selection and model compute using real-time carbon intensity signals. It uses a carbon-elastic curriculum and emission-adaptive knobs, orchestrated by a constrained RL controller. It reduces emissions by ~52% while maintaining near-iso-accuracy.
Unofficial AI-generated summary based on the public title and abstract. Not an official translation.
📝 gxceed 編集解説 — Why this matters
日本のGX文脈において
日本でもAI利用拡大に伴う計算エネルギー消費が課題。本手法はデータセンターのグリーン化やカーボンアウェアネスに示唆を与える。
In the global GX context
As AI compute grows, carbon-aware training frameworks like SUSTAIN-VLM offer practical pathways for reducing operational emissions in cloud and edge deployments.
👥 読者別の含意
🔬研究者:A carbon-elastic framework that dynamically adjusts data and compute for low-carbon VLM fine-tuning.
🏢実務担当者:Can be applied to reduce carbon footprint of AI training pipelines without sacrificing accuracy.
📄 Abstract（原文）
Training and fine-tuning vision–language models (VLMs) consume substantial energy, while grid carbon intensity varies by time and region. We present SUSTAIN-VLM, a carbon-elastic framework that co-adapts (i) data selection via a carbon-elastic curriculum (CEC) and (ii) model compute via emission-adaptive knobs (EACK: resolution, precision, MoE top-k, LoRA rank, checkpointing), orchestrated by a constrained RL controller under real-time carbon signals. SUSTAIN-VLM achieves near-iso-accuracy (–0.7 pts vs. Vanilla) while cutting emissions by ∼52% relative to Vanilla and ∼31% vs. static MoE; it also yields the best emission efficiency (kgCO<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</inf>e per %-accuracy). Results indicate that coordinating what to train and how hard to train across carbon conditions outperforms single-axis strategies such as fixed mixed precision, static MoE, or time-of-day scheduling.
🔗 Provenance — このレコードを発見したソース
- openalex https://doi.org/10.1109/icassp55912.2026.11461984first seen 2026-05-17 05:53:54
🔔 こうした論文の新着を逃したくない方は キーワードアラート に登録（無料・3キーワードまで）。
gxceed は公開メタデータに基づく研究支援データセットです。要約・翻訳・解説は AI 支援で生成されています。 最終的な解釈・検証は利用者が原典資料に基づいて行うことを前提とします。