🏢 Zhejiang University

Adam optimizer achieves near-optimal convergence in non-convex scenarios with unbounded gradients and relaxed noise assumptions, improving its theoretical understanding and practical application.

MoMu-Diffusion: a novel framework that learns long-term motion-music synchronization, generating realistic and beat-matched sequences surpassing existing methods.

Model LEGO (MDA) revolutionizes deep learning by enabling the creation of new models by assembling and disassembling task-aware components from pre-trained models, eliminating the need for retraining.

Researchers taught a large language model (LLM) a three-word ‘Knowledge Graph Language’ (KGL) to improve knowledge graph (KG) completion, drastically reducing errors compared to other methods.

MimicTalk generates realistic, expressive talking videos in minutes using a pre-trained model adapted to individual identities.

MaskFactory generates high-quality synthetic data for dichotomous image segmentation, improving model training efficiency and accuracy.

CATOD framework improves text-to-image generation by actively learning high-quality training data to accurately depict out-of-distribution concepts.

LG-CAV: Train any Concept Activation Vector with Language Guidance, leverages vision-language models to train CAVs without labeled data, achieving superior accuracy and enabling state-of-the-art model…

PFSUM, a novel learning-augmented algorithm, leverages short-term predictions to achieve superior performance in solving the Bahncard problem, outperforming existing methods with improved consistency …

CoupleNet dynamically links protein sequences and structures for improved representations, surpassing state-of-the-art methods in function prediction, particularly for uncommon proteins.

Researchers unveil ‘knowledge circuits’ within LLMs, revealing how knowledge is collaboratively encoded and utilized, leading to improved LLM design and interpretations of model behavior.

InfoRE: A novel method improving large language models’ reasoning by reorganizing information to highlight logical relationships, resulting in a 4% average accuracy boost across various tasks.

A novel algorithm, D-FTBL, achieves improved regret bounds for bandit convex optimization with delayed feedback, tightly matching existing lower bounds in worst-case scenarios.

DiMoP3D: Predicting diverse, physically realistic human motions in 3D scenes by harmonizing stochasticity and determinism.

GDPO: A novel method optimizes graph diffusion models for any objective using reinforcement learning, achieving state-of-the-art performance in diverse graph generation tasks.

FRIEREN: a novel video-to-audio generation network using rectified flow matching achieves state-of-the-art performance by improving audio quality, temporal alignment, and generation efficiency.

FreeLong: Generate high-fidelity long videos without retraining using spectral blending of global and local video features!

FOOGD: A novel federated learning framework that simultaneously tackles out-of-distribution generalization and detection by estimating probability density for reliable global distribution guidance.

FashionR2R leverages diffusion models to realistically translate rendered fashion images into photorealistic counterparts, enhancing realism and preserving fine-grained clothing textures.

Researchers reveal a high risk of training data extraction from molecular pre-trained models, challenging the assumption that model sharing alone adequately protects against data theft.