Image Classification

RTify: A novel framework aligns deep neural networks’ dynamics with human reaction times for improved visual decision-making models.

GLMix: A novel vision backbone efficiently integrates convolutions and multi-head self-attention at different granularities, achieving state-of-the-art performance while addressing scalability issues.

CordsNet: a hybrid CNN-RNN architecture enabling biologically realistic, robust image recognition through continuous-time recurrent dynamics.

ECP-ViT: Real-time Vision Transformer on Mobile Devices via Core-Periphery Attention and Smart Data Layout.

RAMP: A novel training framework significantly boosts DNN robustness against diverse adversarial attacks by mitigating accuracy-robustness tradeoffs and improving generalization.

QuadMamba: A novel vision model leveraging quadtree-based scanning for superior performance in visual tasks, achieving state-of-the-art results with linear-time complexity.

QT-ViT boosts Vision Transformer efficiency by using quadratic Taylor expansion to approximate self-attention, achieving state-of-the-art accuracy and speed.

QKFormer: A groundbreaking spiking transformer achieving 85.65% ImageNet accuracy using a linear-complexity, energy-efficient Q-K attention mechanism.

CutMix and Cutout data augmentation methods provably improve feature learning by enabling the network to learn rarer features and noise vectors more effectively.

Prototypical Hash Encoding (PHE) significantly boosts on-the-fly fine-grained category discovery by using multiple prototypes per category to generate highly discriminative hash codes, thus resolving …

Physics-constrained learning significantly boosts optical neural network accuracy by addressing systematic physical errors, achieving state-of-the-art results on image classification tasks.

BiXT, a novel bi-directional cross-attention Transformer, scales linearly with input size, achieving competitive performance across various tasks by efficiently processing longer sequences.

Boosting vision model training: A new anchored training protocol with a simple regularizer significantly enhances generalization and safety, surpassing standard methods.

Vision Transformers achieve surprisingly high accuracy by transferring only pre-training attention maps, challenging the conventional belief that feature learning is crucial.

CompILer: A novel prompt-based incremental learner mastering state-object compositions without forgetting, achieving state-of-the-art performance.

Self-supervised gradients boost frozen deep learning model performance!

MSVMamba: A novel multi-scale vision model leveraging state-space models, achieves high accuracy in image classification and object detection while maintaining linear complexity, solving the long-rang…

MSPE empowers Vision Transformers to handle any image resolution by cleverly optimizing patch embedding, achieving superior performance on low-resolution images and comparable results on high-resoluti…

MoE Jetpack efficiently transforms readily available dense checkpoints into high-performing MoE models, drastically accelerating convergence and improving accuracy.

Researchers propose a simple yet effective clustering-based adaptive margin loss to mitigate biases inherited by black-box feature extractors in image classification tasks.