Causally Sufficient and Necessary Feature Expansion for Class-Incremental Learning
Causal feature expansion approach for class-incremental learning addressing catastrophic forgetting and feature collisions.
Causal feature expansion approach for class-incremental learning addressing catastrophic forgetting and feature collisions.
Learning netlist representations from imperfect LLM-generated RTL code for hardware design without large labeled datasets.
Transformer model with geological priors for lithology identification from well logs with improved interpretability.
Theoretical bounds for distributed expert selection problem across multiple servers minimizing regret.
Hybrid model combining discrete diffusion and autoregressive language models for improved multi-agent reasoning and planning.
Graph neural network improvement using dual prototype sets to handle global context and noisy local neighborhoods.
Transformer-based signal separator for isolating signals from non-Gaussian interference using learned tokenization.
Game-theoretic approach to multi-agent RL using risk-sensitive equilibrium for robust and efficient agent coordination.
Optimal control formulation for reasoning in language models to enable planning and goal-directed action selection.
Methods for efficient reasoning in Transformers at fixed test-time cost using attention priors and training techniques.
Spiking neural networks for spatiotemporal event-based data classification with improved energy efficiency and temporal decoding.
Method for calibrating measurement error in aggregate outcome variables from surveys and administrative records.
Theoretical analysis connecting training dynamics in Gaussian mixture models to surrogate systems using Gordon comparison theorem.
Reward-Zero: implicit reward mechanism using language embeddings to derive progress signals for reinforcement learning without explicit reward functions.
Graph neural network model for detecting anomalies across multiple domains with adaptive testing-time mechanisms to handle domain shift.
Data-driven framework predicting surface roughness in 3D additive manufacturing using process parameters and surface geometry.
Dataset condensation method for clinical ML models using synthetic data and differential privacy to democratize healthcare AI while protecting patient privacy.
Contrastive learning approach for attributed hypergraph clustering with direct clustering supervision integration.
SPAARS: offline-to-online RL safety method for robotics using abstract exploration and refined action space exploitation.
Improvements to spatial-temporal matching algorithm for GPS trajectory to road network matching in sparse sampling scenarios.
Analysis of MDP design choices impact on sim-to-real transfer in reinforcement learning for industrial process control.
Nonparametric off-policy evaluation method for contextual bandits addressing limitations of inverse probability weighting.
Temporal-conditioned normalizing flows framework for multivariate time series anomaly detection with uncertainty modeling.
XLA-compatible state space model inference implementation enabling O(1) autoregressive caching without NVIDIA hardware dependency.
Constraint-based structure learning for Markov and Bayesian networks with unreliable conditional independence oracles.
Optimal control-theoretic framework for transformer training with structured constraints and McKean-Vlasov dynamics.
Routing mechanism for online continual learning in transformers without forgetting, addressing non-stationary streaming data.
Theoretical analysis of memorization capacity in deep ReLU networks characterized by width and depth parameters.
MM-algorithms for non-negative matrix factorization with Tweedie and Negative Binomial cost functions for unsupervised learning and feature extraction.
Brain network analysis using hierarchical organization learning from fMRI data for brain disorder diagnosis.
Conditional GAN framework for synthesizing subsurface rock formation images from sparse petrography data and well logs.
FreqCycle framework for time series forecasting using multi-scale time-frequency analysis to capture mid to high frequency patterns.
Research on how label and selection bias impact ML classification model evaluation, performance, and mitigation strategies.
Open-source framework evaluating graph neural networks for time series anomaly detection with critical benchmarking.
Empirical study of catastrophic forgetting in LoRA and parameter-efficient fine-tuning methods during sequential learning.
Active learning pipeline for efficiently generating preference data annotations for RLHF-based LLM alignment.
Physics-informed neural operators for accelerating parametric phase-field modeling in materials science.
Federated knowledge distillation approach for AI-native radio access networks in multi-access edge computing systems.
Theoretical generalization bounds for neural oscillators based on second-order ODEs for modeling dynamic systems.
Hybrid quantum-classical framework using quantum circuit born machines for financial volatility forecasting.
Adaptive channel pruning scheme for split learning to reduce communication overhead in distributed training.
Bayesian optimization algorithm for optimizing probability distributions and mixtures on the probability simplex.
In-context reinforcement learning approach that uses high-quality reasoning traces as better demonstrations for improving LLM reasoning.
Lightweight pseudo-projector modification for transformer-based language models to reduce noise sensitivity in hidden representations.
Multi-task multi-fidelity surrogate modeling framework for manufacturing systems using heterogeneous data sources.
Graph attention network model for predicting spectrum demand using geospatial data in wireless networks.
GAST combines gradient-aligned sparse tuning with data-layer selection for parameter-efficient fine-tuning of large language models.
CarbonBench standardizes evaluation of zero-shot spatial transfer learning for upscaling carbon flux measurements across geographic regions.
MSSR is a memory-aware replay strategy for continual LLM fine-tuning that reduces catastrophic forgetting during sequential task learning.
OptEMA optimizer improves exponential moving average with adaptive stepsizes, achieving zero-noise optimality for stochastic optimization.