The paper proposes a memory‑efficient training method for transformers by training a local proxy network for each block to predict the global output via distillation. This allows gradients to be propagated locally while closely approximating the true backpropagation gradients, achieving a better accuracy‑memory trade‑off on CIFAR‑100 vision transformers. The approach also improves fine‑tuning of large language models under tight memory budgets, outperforming other memory‑efficient zeroth‑order algorithms.
arXiv:2506. 14202v4 Announce Type: replace-cross Abstract: End-to-end backpropagation requires storing activations throughout all layers, creating memory bottlenecks that limit model scalability.
arXiv:2605. 04913v4 Announce Type: replace-cross Abstract: LLM post-training typically propagates task gradients through the full depth of the model.
Vision Transformers underperform convolutional networks when training data is scarce, and distilling convolutional inductive biases from a CNN teacher is an effective remedy that leaves the deployed model unchanged. General-purpose feature distillation, however, transfers little in this setting.
AI training's rising resource intensity is straining electricity supplies and carbon budgets, motivating systematic study of memory-efficient training on constrained hardware. We benchmark five gradient optimizers (SGD, Adam, Adagrad, Adadelta, and Conjugate Gradient Descent) under three memory strategies (standard training, gradient checkpointing, and gradient accumulation) across four transformer architectures (ViT, ModernBERT, Llama 3.
arXiv:2610.04867v2 Announce Type: replace-cross Abstract: Greedy block-wise local learning splits a network into gradient-isolated blocks trained by local auxiliary losses, deleting the backward pass...
arXiv:2601. 22580v2 Announce Type: replace-cross Abstract: The success of Large Language Models (LLMs) hinges on the stable training of deep Transformer architectures.
arXiv:2610.20337v1 Announce Type: new Abstract: The success of transformers is attributed to several factors, including their capacity to express formal languages over variable-length inputs, and abi...
arXiv:2608. 08961v1 Announce Type: new Abstract: AI training's rising resource intensity is straining electricity supplies and carbon budgets, motivating systematic study of memory-efficient training on constrained hardware.
arXiv:2106. 06998v5 Announce Type: replace Abstract: Training convolutional neural networks at scale demands substantial memory, largely because intermediate activations must be stored for backpropagation.
GaLe is a memory‑efficient technique that allows pretrained neural networks to run on resource‑constrained devices without retraining. It splits feature maps into a local exact component that keeps fine details and a global approximate component that preserves long‑range dependencies, enabling global operations and attention mechanisms typical of hybrid CNN‑transformer models. On ImageNet, GaLe matches exact‑inference accuracy while delivering up to 65% speedup and 90% RAM reduction on a Cortex‑M33, and it works across classification, detection, and generation tasks.
arXiv:2606. 26538v1 Announce Type: cross Abstract: Deep Transformers are composed of uniformly stacked residual blocks, yet their deepest layers often add little value.
arXiv:2506. 01260v3 Announce Type: replace Abstract: Scaling models has led to significant advancements in deep learning, but training these models in decentralized settings remains challenging due to communication bottlenecks.