Deep Learning

Deep MIL algorithms’ instance-level learnability is theoretically proven, revealing crucial conditions for success and highlighting gaps in existing models.

This paper unveils the Transformer’s approximation power, deriving explicit Jackson-type rates to reveal its strengths and limitations in handling various sequential relationships.

Boosting meta-learning, this paper introduces a novel, flexible approach to create approximately equivariant neural processes that outperform both non-equivariant and strictly equivariant counterparts…

ANT: An adaptive noise schedule automatically determines optimal noise schedules for time series diffusion models, significantly boosting performance across diverse tasks.

A novel multilinear model analytically explains the emergence and scaling laws of skills in the multitask sparse parity problem, accurately predicting skill emergence in neural networks.

Amortized Fourier Neural Operators (AM-FNOs) dramatically improve efficiency in solving PDEs by using neural networks for kernel parameterization, achieving up to 31% better accuracy compared to exist…

ALIDIFF aligns target-aware molecule diffusion models with exact energy optimization, generating molecules with state-of-the-art binding energies and improved properties.

MambaNO: a novel neural operator achieving linear complexity and state-of-the-art accuracy in solving PDEs by cleverly balancing global and local information using an alias-free architecture.

Rate-based backpropagation boosts deep spiking neural network training efficiency by leveraging rate coding, achieving comparable performance to BPTT with reduced complexity.

ADOPT, a novel adaptive gradient method, achieves optimal convergence rates without restrictive assumptions, unlike Adam, significantly improving deep learning optimization.

Multi-Grade Deep Learning (MGDL) conquers spectral bias in deep neural networks by incrementally learning low-frequency components, ultimately capturing high-frequency features through composition.

Adaptive Passive-Aggressive framework with Side information (APAS) significantly boosts online regression accuracy by dynamically adjusting thresholds and integrating side information, leading to supe…

Adaptive Depth Networks with Skippable Sub-Paths: Train once, deploy efficiently! This paper proposes a novel training method to create adaptive-depth networks, enabling on-demand model depth selectio…

Score-based diffusion models are improved by a novel coefficient design, enabling efficient adaptation to unknown low-dimensional data structures and achieving a convergence rate of O(k²/√T).

Ada-MSHyper: A novel adaptive multi-scale hypergraph transformer significantly boosts time series forecasting accuracy by modeling group-wise interactions and handling complex temporal variations.

Slash deep learning’s memory footprint! This paper introduces a novel activation map compression technique via tensor decomposition, significantly boosting on-device training efficiency for edge AI.

Space partitioning dramatically speeds up relative entropy coding (REC) for neural compression, achieving 5-15% better bitrates than previous methods.

Absorb & Escape: a novel post-training sampling method that overcomes single model limitations by combining Autoregressive (AR) and Diffusion Models (DMs), generating high-quality heterogeneous genomi…

ATAC-Diff: A versatile diffusion model for high-quality single-cell ATAC-seq data generation and analysis, surpassing state-of-the-art.

New 2sED measure effectively bounds deep learning model complexity, correlating well with training error and offering efficient computation, particularly for deep models via a layerwise approach.