Deep Learning

Spexphormer achieves significant memory reduction in graph Transformers by leveraging a two-stage training process that leverages attention score consistency across network widths to effectively spars…

This paper provides the first complete error analysis for diffusion models, theoretically justifying optimal training and sampling strategies and design choices for enhanced generative capabilities.

Deep random features enable efficient Euclidean distance compression, offering improved bit storage compared to linear methods for specific parameter ranges, thus significantly advancing high-dimensio…

HyperDM accurately estimates both epistemic and aleatoric uncertainty using a single model, overcoming the computational limitations of existing ensemble methods.

Boosting neural network robustness against weight errors, this research leverages neural tangent kernels to theoretically explain and optimize error-correcting output codes (ECOCs), achieving superior…

Equivariant Neural Diffusion (END) revolutionizes 3D molecule generation with a learnable forward process, achieving state-of-the-art results and enhanced controllability.

Nonlinear spectral filters (NLSFs) enable fully equivariant graph neural networks, improving accuracy and generalization.

Equivariant Blurring Diffusion (EBD) generates 3D molecular conformers hierarchically, first creating coarse-grained fragments then refining atomic details, significantly outperforming existing method…

EnOF-SNN boosts spiking neural network (SNN) accuracy by enhancing output feature representation using a novel knowledge distillation method and ReLU activation, outperforming current state-of-the-art…

Revolutionizing protein mutation effect prediction, this work introduces a retrieval-augmented framework achieving state-of-the-art accuracy by efficiently incorporating similar local structure inform…

Boosting Bayesian deep learning’s diversity and uncertainty quantification, this study proposes hyperspherical energy minimization of CKA to generate diverse and reliable neural network ensembles and …

Train discrete EBMs efficiently with Energy Discrepancy, a novel loss function that eliminates the need for Markov Chain Monte Carlo, using diffusion processes on structured spaces.

Hopfield Boosting, a novel energy-based boosting approach, achieves state-of-the-art OOD detection by leveraging Hopfield energy to sharpen the decision boundary between in-distribution and out-of-dis…

ElasTST: A novel time-series transformer enables robust forecasting across various horizons without per-horizon training, enhancing adaptability and accuracy.

Einspace: A novel neural architecture search space built from fundamental operations, enabling discovery of diverse high-performing network architectures and surpassing existing NAS methods.

EGonc, a novel energy-based open-set node classification method, leverages substitute unknowns and energy scores to achieve superior accuracy and robustness in classifying nodes from known classes whi…

Dynamic rescaling boosts GNN training by controlling layer learning speeds and balancing networks, leading to faster training and improved generalization, especially on heterophilic graphs.

Dynamic Neural Regeneration (DNR) enhances deep learning generalization on small datasets using a data-aware dynamic masking scheme inspired by neurogenesis.

Simulation-free flow generates conditional distributions via dynamic conditional optimal transport.

Dual Cone Gradient Descent (DCGD) enhances Physics-Informed Neural Network (PINN) training by resolving gradient imbalance issues, leading to more accurate and stable solutions for complex partial dif…