Deep Learning

Deep Equilibrium Models excel on imbalanced data due to feature convergence and self-duality properties, unlike explicit models, as shown through Neural Collapse analysis.

Diffusion models’ surprising generalizability stems from an inductive bias towards learning Gaussian data structures, a finding that reshapes our understanding of their training and generalization.

Biologically-inspired ORGANICs neural circuit achieves dynamic divisive normalization, ensuring unconditional stability and seamless backpropagation training for high-dimensional recurrent networks.

UGC: Blazing-fast graph coarsening for big data, preserving key insights across diverse graph types.

TuneTables optimizes PFNs for scalability via context optimization, achieving state-of-the-art performance on large tabular datasets while using fewer parameters and reducing inference time.

Treeffuser: Accurate probabilistic predictions from tabular data using conditional diffusion models and gradient-boosted trees!

Transferable Boltzmann Generators enable efficient, zero-shot sampling of unseen molecular systems’ equilibrium distributions, boosting molecular simulations.

VISPA, a novel BNN training framework using Gaussian variational inference and low-rank SDP, achieves state-of-the-art accuracy on various benchmarks.

Analog in-memory computing (AIMC) training suffers from asymptotic errors due to asymmetric updates. This paper rigorously proves this limitation, proposes a novel discrete-time model to characterize …

N2: A novel dynamic message-passing GNN tackles message-passing bottlenecks and high computational costs by introducing learnable pseudo-nodes and dynamic pathways in a common state space, achieving s…

FKEA: a novel, scalable method for reference-free evaluation of generative models’ diversity using random Fourier features, overcoming computational limitations of existing entropy-based scores.

TinyTTA enables efficient test-time adaptation on memory-constrained edge devices using a novel self-ensemble and early-exit strategy, improving accuracy and reducing memory usage.

TimeXer empowers transformers for superior time series forecasting by cleverly integrating exogenous variables, achieving state-of-the-art results on diverse benchmarks.

Networks trained on continuous sensory data spontaneously develop place cell-like responses, demonstrating that time-encoded experience can create spatial maps in the brain.

This paper introduces a selective G-Bispectrum algorithm, slashing the computational complexity from O(|G|^2) to O(|G|), making G-invariant deep learning faster and more scalable.

Neural networks exhibit ‘Neural Regression Collapse’ (NRC) during training, where feature vectors collapse to subspaces spanned by principal components of features and weights, and the weight vector G…

Poisson Midpoint Method quadratically accelerates Langevin Monte Carlo for diffusion models, achieving high-quality image generation with significantly fewer computations.

I-OBS, a novel family of sparse recovery algorithms leveraging second-order information, achieves faster convergence rates for sparse DNNs, validated by large-scale experiments.

ESCAIP, a novel neural network architecture, dramatically boosts the speed and accuracy of atomic simulations by leveraging attention mechanisms, enabling efficient large-scale modeling across diverse…

Researchers extended implicit bias theory to multiclass classification using a novel framework, proving that gradient descent prefers simple solutions even with complex alternatives.