Image Classification

MambaTree: A novel tree-topology-based state space model surpasses existing methods by dynamically generating input-aware topologies for enhanced long-range dependencies in vision and language.

LookHere: Vision Transformers excel at high-resolution image classification by using 2D attention masks to direct attention heads, improving generalization and extrapolation.

Linearly decompose & recompose Vision Transformers to create diverse-scale models efficiently, reducing computational costs & improving flexibility for various applications.

Meta-learning a hypernetwork on CutMix-augmented data enables data-efficient and precise correction of vision transformer errors by identifying optimal parameters for fine-tuning.

Decoupled Token Embedding for Merging (DTEM) significantly improves Vision Transformer efficiency by using a decoupled embedding module for relaxed token merging, achieving consistent performance gain…

Low-Rank Feature Learning (LRFL) significantly boosts thorax disease classification accuracy by reducing noise and background interference in medical images.

LOFF-TA leverages offline foundation model features and tensor augmentations for efficient, resource-light training, achieving up to 37x faster training and 26x less GPU memory usage.

Learned Bregman divergences significantly improve image corruption robustness in adversarial training.

L-TTA: A lightweight test-time adaptation method using a versatile stem layer minimizes channel-wise uncertainty for rapid and memory-efficient adaptation to new domains.

ProtoViT: a novel interpretable image classification method using Vision Transformers and adaptive prototypes, achieving higher accuracy and providing clear explanations.

This paper introduces L-Reg, a novel logical regularization technique, to improve generalization in visual classification. L-Reg effectively reduces model complexity and improves interpretability by f…

LeTs: Learnable Transformation efficiently initializes variable-sized Vision Transformers by learning adaptable transformations from a compact learngene module, outperforming from-scratch training.

InfiNet achieves state-of-the-art results by enabling feature interaction in an infinite-dimensional space using RBF kernels, surpassing models limited to finite-dimensional interactions.

This research leverages causal intervention to identify and utilize genuine label correlations in multi-label image recognition, mitigating contextual bias for improved accuracy.

Boost DNN robustness to corruptions without sacrificing clean image accuracy using Data Augmentation via Multiplicative Perturbations (DAMP)!

HydraViT: Stacking attention heads creates a scalable Vision Transformer, adapting to diverse hardware by dynamically selecting subnetworks during inference, improving accuracy and efficiency.

Hierarchical Selective Classification (HSC) improves deep learning model reliability for risk-sensitive tasks by leveraging hierarchical class relationships to provide more informative predictions eve…

Happy: a novel debiased learning framework, excels at continually discovering new categories from unlabeled data while retaining knowledge of previously learned ones, overcoming existing bias issues a…

Guaranteed Manifold Clustering: Novel method provides geometric conditions ensuring accurate data grouping from nonlinear manifolds, showing competitive performance on CIFAR datasets.

FlipClass dynamically updates the teacher model in a teacher-student framework to align with the student’s attention, resolving learning inconsistencies and significantly improving generalized categor…