Deep Learning

Accelerate Bayesian inference in linear mixed-effects models by efficiently marginalizing random effects using fast linear algebra, enabling faster and more accurate posterior estimations.

Researchers devised a novel method to accelerate neural network training by guiding the optimization process toward a Simplex Equiangular Tight Frame, exploiting the Neural Collapse phenomenon to enha…

Group-and-Shuffle (GS) matrices enable efficient structured orthogonal parametrization, improving parameter and computational efficiency in orthogonal fine-tuning for deep learning.

Deep neural networks surprisingly exhibit universal convergence in input salience, aligning more closely as model capacity increases, revealing valuable insights into model behavior and improving deep…

GraphMETRO tackles complex graph distribution shifts by using a Mixture-of-Experts model to decompose shifts into interpretable components, achieving state-of-the-art results.

Boost GNN performance and overcome oversmoothing with Cluster-Normalize-Activate (CNA) modules: a simple yet highly effective plug-and-play solution!

GNeuralFlow unveils systemic interactions in irregularly sampled time series by learning a directed acyclic graph representing conditional dependencies, achieving superior performance in classificatio…

GRAPHEDX, a novel neural network, accurately estimates graph edit distance with varying operation costs, outperforming existing methods.

GRDL: a novel graph classification method boasting 10x speed improvement over competitors, achieved by treating node embeddings as distributions and avoiding global pooling.

GRANOLA: A novel graph-adaptive normalization layer significantly boosts GNN performance by dynamically adjusting node features based on the input graph’s unique structure.

ACS: Automatic Cyclical Scheduling revolutionizes gradient-based discrete sampling by intelligently switching between exploration and exploitation phases to efficiently navigate complex multimodal dis…

Gradient Rewiring (GRE) improves editable GNN training by addressing gradient inconsistencies, preserving training node performance while correcting target node errors.

Train factorized neural network layers efficiently with Geometry-aware training in Tucker format (TDLRT)!

RNNs represent naturalistic objects in WM using chronological subspaces, defying traditional slot models; object features are less orthogonalized in RNNs vs. perceptual space.

Cross-geometry learning using knowledge distillation significantly improves GNN performance by leveraging both Euclidean and hyperbolic geometric properties of graph data.

GeoTDM: First diffusion model generating realistic 3D geometric trajectories, capturing complex spatial interactions and temporal correspondence, significantly improving generation quality.

GENOT: a flexible neural optimal transport framework for single-cell genomics, enabling stochastic map learning with any cost function, handling unbalanced data, and tackling complex (Fused) Gromov-Wa…

Genetic-guided GFlowNets revolutionize sample-efficient molecular optimization by smartly integrating genetic algorithms into GFlowNets training, achieving state-of-the-art performance with substantia…

MDGEN: Generative modeling unlocks MD data for diverse tasks, achieving significant speedups via flexible multi-task surrogate models.

QGCNs generalize CNNs to graph data via learnable neighborhood quantization, achieving state-of-the-art performance on graph datasets.