AI Paper Atlas

Research map / Architectures and efficiency / RNNs and normalization

Efficient Resource-Constrained Training of Transformers via Subspace Optimization

Le-Trung Nguyen, Enzo Tartaglione, Van-Tam Nguyen

ICLR 2026 · Architectures and efficiency › RNNs and normalization · approach: Scalable approximation and equivariant compute · 1 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
use warm-started subspace iteration (WSI) instead of full SVD at every iteration (ablation)WASI weight decomposition · ViT Petstraining FLOPs at equal accuracy: 1.36x fewer FLOPs; about 35% higher accuracy at equal FLOPsImproved
raise explained variance threshold from 0.4 to 0.9 (ablation)WASI · ViT CIFAR-10 (all linear layers)accuracy (%): 68.99 -> 96.24Improved
raise explained variance threshold from 0.4 to 0.9 (ablation)WASI · ViT CIFAR-10 (all linear layers)training memory (MB): 39.39 -> 179.61Hurt
apply WASI at epsilon 0.9 instead of vanilla trainingViT · ViT CIFAR-10 (all linear layers)accuracy (%): 97.32 -> 96.24No clear effect
apply WASI at epsilon 0.9 instead of vanilla trainingViT · ViT CIFAR-10 (all linear layers)training memory (MB): 2349.00 -> 179.61Improved
apply WASI at epsilon 0.9 instead of vanilla trainingViT · ViT CIFAR-10 on Raspberry Pi 5training time (s): 23.87 -> 16.57Improved
compress weights as well as activations (WASI vs ASI) at epsilon 0.9ASI · ViT CIFAR-10 on Raspberry Pi 5training time (s): 25.52 -> 16.57Improved
compress weights as well as activations (WASI vs ASI) at epsilon 0.4ASI · ViT CIFAR-10 on Raspberry Pi 5inference time (s): 7.69 -> 3.15Improved
apply WSI to last 1 to 4 convolutional layers at small epsilonMCUNet · Petsweight memory and accuracyNo clear effect
apply WSI to convolutional layers at epsilon 0.9MCUNet · Petsweight memoryHurt
apply WASI at epsilon 0.9 instead of vanilla trainingSwinT · multiple datasetsmemory and FLOPs: memory by up to 62x; FLOPs by 1.5xImproved
apply WASI instead of vanilla trainingTinyLlama · BoolQactivation and weight memory: 953.86x and 30.12x reductionImproved

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy