Spotlight Others

SSCNN, a novel decomposition-based model, achieves superior long-term time series forecasting accuracy using 99% fewer parameters than existing methods, proving that bigger isn’t always better.

Parameter-Inverted Image Pyramid Networks (PIIP) boost image pyramid efficiency by using smaller models for higher-resolution images and larger models for lower-resolution ones, achieving superior per…

Researchers visualized shared visual features driving responses of body-selective neurons to non-body objects, revealing object parts resembling macaque body parts, thus explaining neural preferences.

PACE marries parameter-efficient fine-tuning with consistency regularization to significantly boost model generalization.

Researchers developed a new evaluation metric, AUGRC, for selective classification systems that overcomes the limitations of existing metrics by providing a more holistic and interpretable assessment …

Deep learning optimally learns holomorphic operators between Banach spaces, achieving near-optimal generalization bounds with problem-agnostic DNN architectures.

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

Unlocking superior discriminative features from diffusion models, this research reveals key activation properties for effective feature selection, surpassing state-of-the-art methods.

New neural operator, Nonlocal Attention Operator (NAO), simultaneously learns forward and inverse physical models, improving interpretability and generalizability for physics discovery.

Neural receptive fields’ localization emerges from nonlinear learning dynamics driven by naturalistic data’s higher-order statistics, not just sparsity.

RUM neural network, a novel non-convolutional GNN, overcomes limitations of conventional convolution-based models by using RNNs to merge topological and semantic features along random walks, achieving…

Data-driven approach corrects confidence intervals in high-dimensional learning, improving accuracy for various models and bridging theory and practice.

New non-asymptotic approximation error bounds show that parameterized quantum circuits can efficiently approximate complex functions, potentially surpassing classical neural networks.

Crowd wisdom solves noisy label learning!

Neural Krylov Iteration (NeurKItt) accelerates linear system solving by using a neural operator to predict invariant subspaces, drastically reducing iteration counts and computation time.

Neural Assets enables intuitive 3D multi-object scene editing via image diffusion models by using per-object representations to control individual object poses, achieving state-of-the-art results.

Deep learning’s mysteries surrounding sharpness regularization are solved by uncovering the crucial role of the neglected Hessian component, the Nonlinear Modeling Error (NME).

A novel diffusion model reconstructs multimodal shape distributions from shading, mirroring human multistable perception.

MultiOOD benchmark and novel A2D & NP-Mix algorithms drastically improve multimodal out-of-distribution detection.

Boosting vision-language models: Multilingual data improves performance on English-centric benchmarks.