Core Concepts
Dynamic Tuning (DyT) improves parameter and inference efficiency for ViT adaptation.
Abstract
Existing PEFT methods focus on parameter efficiency, but DyT addresses both parameter and inference efficiency.
DyT introduces a token dispatcher to dynamically select tokens for processing, reducing redundant computation during inference.
Model variants explore different strategies for skipping tokens in transformer blocks.
MoE-adapter enhances adapter capability without increasing computational cost significantly.
Results show DyT's effectiveness across various tasks and datasets, achieving comparable performance with reduced FLOPs.
Introduction:
Vision Transformers (ViTs) require efficient adaptation methods to reduce computational costs.
Parameter-efficient fine-tuning:
PEFT methods reduce tunable model parameters while maintaining fine-tuning accuracy.
Inference efficiency remains unexplored in existing methods.
Dynamic Tuning Approach:
DyT introduces a token dispatcher to selectively activate tokens in transformer blocks.
Different model variants explore strategies for skipping tokens before different blocks.
MoE-adapters enhance adapter capability without increasing computational cost significantly.
Experimental Results:
DyT achieves comparable performance with reduced FLOPs across various tasks and datasets.
Stats
PEFT方法は、チューニング問題を解決するために提案されました。
PEFT方法は、学習可能なモデルパラメータを減らすことで精度を維持します。
PEFT方法は、推論効率の向上に焦点を当てています。