↓Skip to main content

🏢 University of Michigan

From Unstructured Data to In-Context Learning: Exploring What Tasks Can Be Learned and When
·1923 words·10 mins· loading · loading
Natural Language Processing Large Language Models 🏢 University of Michigan
LLMs’ in-context learning surprisingly arises from simple co-occurrence patterns in unstructured data, but positional information is key for complex tasks; ICL fails when patterns are unseen or fixed.
Fine-grained Analysis of In-context Linear Estimation: Data, Architecture, and Beyond
·1351 words·7 mins· loading · loading
Natural Language Processing Large Language Models 🏢 University of Michigan
Researchers crack the code of in-context learning in Transformers, revealing how architecture, low-rank parameters, and data correlations influence model optimization and generalization.
FIDE: Frequency-Inflated Conditional Diffusion Model for Extreme-Aware Time Series Generation
·2091 words·10 mins· loading · loading
Machine Learning Generative Learning 🏢 University of Michigan
FIDE, a novel conditional diffusion model, accurately generates time series by inflating high-frequency components, preserving extreme value distributions.
Exploring the trade-off between deep-learning and explainable models for brain-machine interfaces
·2641 words·13 mins· loading · loading
AI Generated AI Applications Healthcare 🏢 University of Michigan
KalmanNet, a novel BMI decoder, achieves state-of-the-art performance by integrating recurrent neural networks into Kalman filtering, balancing accuracy and explainability.
Exploring Low-Dimensional Subspace in Diffusion Models for Controllable Image Editing
·2111 words·10 mins· loading · loading
Computer Vision Image Generation 🏢 University of Michigan
LOCO Edit achieves precise, localized image editing in diffusion models via a single-step, training-free method leveraging low-dimensional semantic subspaces.
Efficient multi-prompt evaluation of LLMs
·2504 words·12 mins· loading · loading
Natural Language Processing Large Language Models 🏢 University of Michigan
PromptEval efficiently estimates LLM performance across many prompts, providing robust performance metrics and enabling reliable LLM comparisons.
Efficient Contextual LLM Cascades through Budget-Constrained Policy Learning
·3825 words·18 mins· loading · loading
Natural Language Processing Large Language Models 🏢 University of Michigan
TREACLE: a reinforcement learning policy efficiently selects LLMs and prompts, achieving up to 85% cost savings while maintaining high accuracy in answering reasoning questions.
Distributionally Robust Performative Prediction
·2341 words·11 mins· loading · loading
AI Generated AI Theory Optimization 🏢 University of Michigan
This research introduces distributionally robust performative prediction, offering a new solution concept (DRPO) that minimizes performative risk even with misspecified distribution maps, ensuring rob…
Distributed Least Squares in Small Space via Sketching and Bias Reduction
·1322 words·7 mins· loading · loading
Machine Learning Optimization 🏢 University of Michigan
Researchers developed a novel sparse sketching method for distributed least squares regression, achieving near-unbiased estimates with optimal space and time complexity.
DiffusionPDE: Generative PDE-Solving under Partial Observation
·3911 words·19 mins· loading · loading
Machine Learning Deep Learning 🏢 University of Michigan
DiffusionPDE uses generative diffusion models to solve PDEs accurately, even with highly incomplete observations, outperforming state-of-the-art methods.
DiffusionBlend: Learning 3D Image Prior through Position-aware Diffusion Score Blending for 3D Computed Tomography Reconstruction
·2570 words·13 mins· loading · loading
Computer Vision 3D Vision 🏢 University of Michigan
DiffusionBlend++ learns a 3D image prior via position-aware diffusion score blending, achieving state-of-the-art 3D CT reconstruction with superior efficiency.
CONTRAST: Continual Multi-source Adaptation to Dynamic Distributions
·2633 words·13 mins· loading · loading
Machine Learning Domain Adaptation 🏢 University of Michigan
CONTRAST efficiently adapts multiple source models to dynamic data distributions by optimally weighting models and selectively updating only the most relevant ones, achieving robust performance withou…
BLAST: Block-Level Adaptive Structured Matrices for Efficient Deep Neural Network Inference
·3033 words·15 mins· loading · loading
Computer Vision Image Generation 🏢 University of Michigan
BLAST matrix learns efficient weight structures for faster deep learning inference, achieving significant compression and performance gains on various models.
AutoGuide: Automated Generation and Selection of Context-Aware Guidelines for Large Language Model Agents
·2274 words·11 mins· loading · loading
Natural Language Processing Large Language Models 🏢 University of Michigan
AutoGuide: Automated generation of context-aware guidelines significantly improves LLM agent performance in unfamiliar domains.
Abrupt Learning in Transformers: A Case Study on Matrix Completion
·5285 words·25 mins· loading · loading
AI Generated Natural Language Processing Large Language Models 🏢 University of Michigan
Transformers exhibit abrupt learning: training loss plateaus, then suddenly drops. This study uses matrix completion to demonstrate this phenomenon, providing insights into the model’s algorithmic sh…