Image Classification

Hierarchical concept bottleneck models boost interpretability and accuracy in visual classification by uncovering both high-level and low-level concepts.

InLine attention, a novel method, bridges the performance gap between softmax and linear attention by incorporating injectivity and local modeling, achieving superior performance while maintaining lin…

Researchers discover that existing backdoor defenses leave vulnerabilities, allowing for easy re-activation of backdoors through subtle trigger manipulation.

Boolean Logic Deep Learning (BOLD) revolutionizes deep learning by enabling training with Boolean weights and activations, achieving state-of-the-art accuracy with drastically reduced energy consumpti…

Biologically-inspired AI model integrates learning & visual guidance via a novel ‘Counter-Hebb’ learning mechanism, achieving competitive performance on multi-task learning benchmarks.

APM: Asynchronous Perception Machine, a computationally-efficient architecture for test-time training (TTT), processes image patches asynchronously, encoding semantic awareness without pre-training, a…

Focusing on nuclear morphology improves out-of-domain generalization in cancer classification from histopathology images by leveraging nuclear segmentation masks during training.

Large-scale dataset distillation can be achieved with significantly less soft labels by using class-wise supervision during image synthesis, enabling simple random label pruning and enhancing model ac…

SaSPA, a novel data augmentation method, boosts fine-grained visual classification accuracy by generating diverse, class-consistent synthetic images using structural and subject-preserving techniques.

Adaptive Randomized Smoothing certifies deep learning model predictions against adversarial attacks by cleverly combining randomized smoothing with adaptive, multi-step input masking for improved accu…

Boosting Vision Transformer robustness against attacks & noisy data, AdaNCA uses Neural Cellular Automata as plug-and-play adaptors between ViT layers, achieving significant accuracy improvement with …

KnoBo enhances deep learning models for medical image analysis by incorporating knowledge priors from medical textbooks, boosting out-of-domain performance by up to 32.4%.

Soft labels, not sophisticated data synthesis, are the key to successful dataset distillation, significantly improving data-efficient learning and challenging existing methods.