Image Classification

This research dramatically accelerates neighborhood attention, a cost-effective self-attention mechanism, through novel GEMM-based and fused kernel implementations, boosting performance by up to 1759%…

F-OAL: Forward-only Online Analytic Learning achieves high accuracy and low memory usage in online class incremental learning by using a frozen encoder and recursive least squares to update a linear …

This paper introduces a novel token pruning method for vision state space models, achieving significant computational reduction with minimal performance impact, addressing the limitations of directly …

DiChaViT boosts channel-adaptive vision transformers by enhancing feature diversity, yielding a 1.5-5% accuracy gain over state-of-the-art MCI models on diverse datasets.

EMR-MERGING: A tuning-free model merging technique achieves high performance by electing a unified model and generating lightweight task-specific modulators, eliminating the need for additional data …

Elucidating Dataset Condensation (EDC) achieves state-of-the-art accuracy in dataset condensation by implementing soft category-aware matching and a smoothing learning rate schedule, improving model t…

Sort & Search: 1000x faster lifelong model evaluation!

Boosting Vision Transformer adaptation! Householder Transformation-based Adaptor (HTA) outperforms existing methods by dynamically adjusting adaptation matrix ranks across layers, improving efficiency…

Dynamic Tuning (DyT) significantly boosts Vision Transformer (ViT) adaptation by dynamically skipping less important tokens during inference, achieving superior performance with 71% fewer FLOPs than e…

Boosting dataset distillation, a new method, Diversity-Driven Synthesis, uses directed weight adjustment to create diverse, representative synthetic datasets, improving model performance while reducin…

DistDiff, a training-free data expansion framework, leverages distribution-aware diffusion models to generate high-fidelity, diverse samples that enhance downstream model performance.

DiffuLT uses a novel diffusion model to generate balanced training data from imbalanced datasets, achieving state-of-the-art results in long-tailed image recognition without external knowledge.

DEX boosts CNN accuracy on tiny AI accelerators by 3.5%p, utilizing unused memory and processors to extend input channels without increasing latency.

DeSparsify: A stealthy adversarial attack exhausts vision transformer resources by exploiting token sparsification mechanisms’ dynamic nature, highlighting the need for improved resource management i…

DEPrune: A novel GPU-optimized pruning method for depthwise separable convolutions, achieving up to 3.74x speedup on EfficientNet-B0 with no accuracy loss!

DenoiseRep: A novel denoising model enhances feature discrimination in computer vision tasks by integrating feature extraction and denoising within a single backbone, achieving impressive improvements…

Vision’s Mamba model demystified: Researchers unveil its surprising link to linear attention, improving efficiency and accuracy through design enhancements.

Improved Kullback-Leibler (IKL) divergence loss achieves state-of-the-art adversarial robustness and competitive knowledge distillation performance by addressing KL loss’s limitations.

CUFIT: a novel curriculum fine-tuning paradigm significantly improves medical image classification accuracy despite noisy labels by leveraging pre-trained Vision Foundation Models.

Researchers developed semantics-aware adversarial examples using a probabilistic approach, achieving higher success rates in bypassing defenses while remaining undetectable to humans.