PAC-bayesian learning
PulseAugur coverage of PAC-bayesian learning — every cluster mentioning PAC-bayesian learning across labs, papers, and developer communities, ranked by signal.
4 day(s) with sentiment data
-
New theory unifies generalization, validation, and information fusion in ML
A new theoretical framework, gamma-CUBV, has been proposed to unify generalization, validation, and information fusion in machine learning. This framework generalizes Cross Upper-Bound Validation (CUBV) by controlling t…
-
New Algorithm Derandomizes Stochastic Majority Votes Using PAC-Bayesian Theory
A new paper introduces a framework to derandomize stochastic majority votes, a key component in ensemble machine learning methods. By applying disintegrated PAC-Bayesian theory, the research transforms existing stochast…
-
New nonlinear dimensionality reduction techniques enhance Bayesian optimization
Researchers have developed new nonlinear dimensionality reduction techniques for Bayesian optimization, a method used for efficient global optimization of expensive black-box functions. The proposed approach, SDR-LSBO, …
-
New PAC-Bayesian Framework Quantifies Value of Privileged Information in ML
Researchers have developed a new PAC-Bayesian framework to quantify the value of privileged information (PI) in machine learning. This approach offers an algorithm-agnostic method to estimate the potential knowledge tra…
-
New PAC-Bayesian Bounds Developed for Partially Observed LTI Systems
Researchers have developed new PAC-Bayesian error bounds for partially observed stochastic linear time-invariant state-space systems that include inputs and sub-Gaussian noise. These bounds connect the expected predicti…
-
New FragileFlow Method Boosts Foundation Model Robustness
Researchers have introduced FragileFlow, a novel plug-in regularizer designed to enhance the robustness of foundation models, including LLMs and Vision-Language Models. This method addresses a failure mode where predict…
-
New PAC-Bayesian Framework Enhances Time Series VAE Guarantees
Researchers have developed a new PAC-Bayesian framework to provide generalization guarantees for Variational Autoencoders (VAEs) when applied to time series data. This framework extends existing PAC-Bayesian guarantees …
-
New paper explores information flow in martingales, unifying concentration inequalities
A new paper published on arXiv details advancements in understanding information flow within the path space of nonnegative martingales. The research introduces exact variational identities that apply even at arbitrary r…
-
New PAC-Bayes Theory Focuses on Behavioral Equivalence and Z-Information
A new research paper introduces PAC-Bayes theory that extends beyond parameter space, focusing on behavioral equivalence and Z-information. The study formalizes behavioral equivalence using a measurable behavior map and…
-
New sampling method boosts long-tailed learning accuracy
Researchers have developed a new method called Sharpness-Guided Equilibrium Sampling (SGS) to improve the performance of models trained on long-tailed datasets. SGS dynamically adjusts the sampling probability of data p…
-
MasFACT framework tackles topology forgetting in multi-agent LLM systems
Researchers have introduced MasFACT, a novel framework designed to address "topology forgetting" in continual multi-agent systems (MAS) powered by large language models. This issue arises when adapting to new tasks caus…
-
New PAC-Bayesian framework explains adversarial training overfitting
Researchers have developed a new PAC-Bayesian analytical framework to understand the phenomenon of robust overfitting in adversarial training. By modeling adversarial training with momentum SGD as a discrete-time dynami…
-
Entanglement, not parameters, governs quantum policy generalization
A new research paper proposes that entanglement, rather than the number of parameters, is the key factor determining generalization in quantum reinforcement learning policies. The study introduces a PAC-Bayesian framewo…
-
New PAC-Bayesian method offers control certification for quadratic systems
Researchers have developed a new method using PAC-Bayesian bounds to certify quadratic closed-loop control systems. This approach addresses challenges with unbounded and non-Lipschitz loss functions by employing System …
-
New multi-distribution Rényi divergences characterized by researchers · 2 sources tracked
Researchers have characterized a new family of multi-distribution generalizations of Rényi divergences, which are crucial for comparing multiple probability distributions simultaneously. This new family, termed multi-wa…
-
New algebraic identity unifies information theory results
A new paper introduces a unified algebraic identity that connects various information-theoretic variational results. This identity generalizes classical formulas for entropy and divergence to multiple priors and holds f…
-
Paper: LLMs face fundamental limits as general-purpose solvers via prompting
A new paper argues that large language models (LLMs) are not truly general-purpose solvers due to fundamental constraints of prompt-based communication. The research suggests that language itself is a limited channel fo…
-
New paper questions LLM general-purpose learning limits due to language constraints
A new arXiv paper argues that large language models (LLMs) are not truly general-purpose learners due to fundamental constraints imposed by natural language as an interface. The research introduces the concepts of an "e…
-
New PAC-Bayes Derandomization Method for Smooth Loss Functions
Researchers have developed a new method for derandomizing PAC-Bayes generalization bounds, specifically for smooth loss functions. This approach aims to create high-probability bounds for deterministic predictors by lev…
-
New method predicts data distributions under drift and corruption
Researchers have developed a novel online learning method for predicting full data-generating distributions in non-stationary data streams, even when subjected to drift and adversarial corruption. The approach utilizes …