Deep Learning

SµPar stabilizes sparse neural network training, slashing tuning costs and boosting performance, especially at high sparsity levels, via a novel parameterization technique.

A new learnable prior for compressive sensing solves the inverse problem using only a few corrupted data samples, enabling sparse signal recovery without ground-truth information and uncertainty quant…

Sourcerer: A novel sample-based method for maximum entropy source distribution estimation, resolving ill-posedness while maintaining simulation accuracy.

Soft ascent-descent (SoftAD) improves test accuracy and generalization by softening the flooding method, offering competitive accuracy with reduced loss and model complexity.

SLU: a novel, low-memory uncertainty score for neural networks, achieves logarithmic memory scaling with model parameters, providing well-calibrated uncertainties and outperforming existing methods.

Train Neural ODEs without simulations, achieving high performance on regression and classification by using flow matching in the embedding space of data pairs.

Sigmoid gating significantly boosts sample efficiency in Mixture of Experts models compared to softmax gating, offering faster convergence rates for various expert functions.

SharpBalance, a novel training approach, effectively improves deep ensemble performance by addressing the sharpness-diversity trade-off, leading to significant improvements in both in-distribution and…

Set-based Neural Network Encoder (SNE) efficiently encodes neural network weights for property prediction, eliminating the need for architecture-specific models and improving generalization across dat…

SequentialAttention++ unites differentiable pruning with combinatorial optimization for efficient and accurate neural network block sparsification, achieving state-of-the-art results.

This paper introduces a novel, label-free method for detecting harmful distribution shifts in machine learning models deployed in production environments, leveraging a proxy error derived from an erro…

Self-Refining Diffusion Samplers (SRDS) dramatically speeds up diffusion model sampling by leveraging Parareal iterations for parallel-in-time computation, maintaining high-quality outputs.

Revolutionizing large neural networks, this paper introduces a continuous parameterization of structured matrices, discovering that full-rank structures without parameter sharing achieve optimal scali…

Researchers developed a novel algorithm to automatically find optimal schedules for denoising diffusion models (DDMs), significantly improving sample quality and efficiency without manual parameter tu…

FuncMol: A new neural field model generates 3D molecules efficiently, outperforming existing methods by achieving an order of magnitude faster sampling speed.

TRODO: a novel trojan detection method using out-of-distribution samples, effectively identifies trojaned classifiers even against adversarial attacks and with limited data.

Stormer, a simple transformer model, achieves state-of-the-art medium-range weather forecasting accuracy by using weather-specific embedding, randomized dynamics forecasting, and a pressure-weighted l…

Boost machine learning with surrogate data! A novel weighted ERM method effectively integrates surrogate data, significantly reducing test errors even with unrelated data, guided by a predictable sca…

Unlocking the potential of deep learning for time series forecasting: this study reveals a scaling law influenced by dataset size, model complexity, and the crucial look-back horizon, leading to impro…

Scale-invariant optimal subsampling tackles computational challenges in analyzing massive rare-events data with sparse models, enhancing parameter estimation and variable selection without being affec…