Spotlight Others

Deep learning privacy is enhanced by a new membership inference attack using input loss curvature, exceeding existing methods, especially on large datasets.

CooHOI: A two-phase learning framework enables physically simulated characters to perform cooperative object transportation tasks naturally and efficiently, overcoming the limitations of existing meth…

CGFormer: a novel voxel transformer boosting semantic scene completion accuracy by using context-aware queries and 3D deformable attention, outperforming existing methods on SemanticKITTI and SSCBench…

C-JEPA boosts self-supervised visual learning by integrating contrastive learning with a joint-embedding predictive architecture, enhancing stability and representation quality.

Conditioning infinite-dimensional nonlinear diffusion processes is made possible, enabling analysis of complex data like organism shapes in evolutionary biology.

Sparse Differentiable Tree Machine (sDTM) improves compositional generalization in neural networks by efficiently representing tree structures in vector space, enabling simultaneous symbolic and neura…

Cluster-wise Graph Transformer (Cluster-GT) improves graph learning by using a novel Node-to-Cluster Attention mechanism that leverages multiple kernel learning to capture node and cluster-level infor…

Boosting multimodal contrastive learning, this research introduces negCLIPLoss and NormSim, novel data selection methods surpassing existing techniques by improving data quality and task relevance. Th…

scCello: A Cell Ontology-Guided Transcriptome Foundation Model improves single-cell RNA sequencing analysis by incorporating cell lineage information, significantly boosting accuracy and generalizabil…

GF-SAM: A novel graph-based few-shot semantic segmentation method leverages SAM’s power efficiently via positive-negative prompt alignment and mask clustering for superior accuracy and speed.

BricksRL: A low-cost, open-source platform democratizes robotics and reinforcement learning research using LEGO, enabling accessible real-world experiments.

Controllable long-tailed learning achieved via hypernetwork-generated diverse experts, adapting to user preferences and distribution shifts.

Brain-JEPA: a novel brain dynamics foundation model leverages fMRI data via innovative gradient positioning and spatiotemporal masking to achieve state-of-the-art performance in diverse brain activity…

BPQP: A new differentiable convex optimization framework accelerates end-to-end learning by an order of magnitude, achieving significant efficiency gains over existing methods.

TransCLIP significantly boosts vision-language model accuracy by efficiently integrating transduction, a powerful learning paradigm that leverages the structure of unlabeled data.

BMRS: Bayesian Model Reduction for Structured Pruning offers a principled, threshold-free approach to neural network compression, achieving high accuracy and competitive efficiency.

BACKTIME unveils effective backdoor attacks on multivariate time series forecasting, highlighting vulnerabilities and offering novel defense strategies.

Autoregressive image generation is revolutionized by eliminating vector quantization, achieving strong results with increased speed using a novel diffusion procedure.

AI autonomously designs highly effective hybrid digital twins by combining neural networks and mechanistic models, significantly advancing digital twin technology.

New framework, MAS, effectively mitigates stereotypes in text-to-image generation by aligning the probability distribution of generated images to stereotype-free distributions.