🏢 Georgia Institute of Technology

Zeroth-Order Diffusion Monte Carlo (ZOD-MC) efficiently samples from non-log-concave distributions using only zeroth-order queries, overcoming metastability issues and outperforming state-of-the-art s…

Vaccine: a novel technique safeguards LLMs against harmful fine-tuning attacks by creating invariant hidden embeddings.

Deep learning scaling laws are explained by novel approximation and estimation theories for transformers on low-dimensional data, resolving discrepancies between theory and practice.

DANCE, a novel semi-supervised framework, efficiently transfers cell types across multi-omic single-cell data even with limited labeled samples, outperforming current state-of-the-art methods.

Selective Projection Decay (SPD) enhances robust fine-tuning of foundation models by selectively applying weight decay, improving generalization and out-of-distribution robustness.

Accelerated Lie group optimization achieved via a novel momentum algorithm (Lie NAG-SC) with proven convergence rates, surpassing existing methods in efficiency.

This paper proves Nesterov’s Accelerated Gradient achieves faster convergence for rectangular matrix factorization and linear neural networks, using a novel unbalanced initialization.

New analysis reveals how reweighted least-squares algorithms for linear diagonal networks achieve favorable performance in high-dimensional settings, improving upon existing theoretical guarantees and…

ORI: a novel online relational inference framework efficiently identifies hidden interaction graphs in evolving multi-agent systems using streaming data and real-time adaptation.

Lisa: a novel lazy safety alignment method safeguards LLMs against harmful fine-tuning attacks by introducing a proximal term to constrain model drift, significantly improving alignment performance.

ELSA: a new model that learns spatially aware language and audio embeddings, achieving state-of-the-art performance in semantic retrieval and 3D sound source localization.

Large Pre-trained Time-series Models (LPTM) achieves superior forecasting and time-series classification results using a novel adaptive segmentation method, requiring up to 40% less data and 50% less …

Langevin unlearning offers a novel, privacy-preserving machine unlearning framework based on noisy gradient descent, handling both convex and non-convex problems efficiently.

HYDRA, a novel model factorization framework, significantly improves black-box LLM personalization by capturing both user-specific behavior and shared knowledge, achieving a 9.01% average relative imp…

Adversarial training achieves minimax-optimal prediction error in high-dimensional linear regression under l∞-perturbation, improving upon existing methods.

BeNeDiff uses generative diffusion models to disentangle and interpret neural dynamics linked to specific behaviors, providing interpretable quantifications of behavior in multi-brain region datasets.

DP-Attacker unveils diffusion-based policy vulnerabilities by crafting effective adversarial attacks, significantly impacting robot safety and paving the way for more robust AI.

This paper introduces a novel differentially private graph diffusion framework ensuring edge-level privacy, significantly improving utility-privacy trade-offs for personalized PageRank computation.

Derivative-enhanced DeepONets boost PDE solution accuracy and derivative approximation, particularly valuable with limited training data.

This paper introduces a novel machine unlearning method using projected noisy stochastic gradient descent, providing the first approximate unlearning guarantee under convexity, significantly improving…