Fisher information
PulseAugur coverage of Fisher information — every cluster mentioning Fisher information across labs, papers, and developer communities, ranked by signal.
5 day(s) with sentiment data
-
New method optimizes label acquisition for multiclass classification risk
Researchers have developed a novel method for optimizing the allocation of limited labeling budgets in multiclass classification tasks. The approach combines Fisher information from acquired labels with the local geomet…
-
New research establishes tight bounds for quantum state tomography with limited entanglement
Two new research papers, one from arXiv and another highlighted by Hugging Face, have established tight lower bounds for quantum state tomography. These studies focus on scenarios where measurements are limited to actin…
-
New Hessian Analysis Method Aids Compression of Billion-Parameter Models
Researchers have developed a novel scalable Kronecker-based approximation to analyze the Hessian of billion-parameter language models. This method efficiently captures cross-layer interactions without needing to store t…
-
New method tackles semi-supervised classification with informative missing labels
Researchers have developed a new semi-supervised classification method for data with missing labels, specifically addressing scenarios where the probability of a missing label is dependent on the observed features. This…
-
New theory quantifies convergence for Langevin-regularized SVGD
This paper introduces a new theoretical framework for understanding Langevin-regularized Stein Variational Gradient Descent (SVGD). The research establishes quantitative convergence guarantees to the target distribution…
-
New method generates synthetic data for conversational recommender systems
Researchers have developed a method for bootstrapping conversational recommender systems (CRS) without requiring domain-specific dialogue data. This approach generates synthetic conversational supervision using non-conv…
-
New research explores uncertainty-dependent missing labels for classification
A new research paper explores a semi-supervised classification method where the probability of a label being missing is dependent on the observed features and the classifier's uncertainty. This approach treats missingne…
-
New defense framework GraphRP protects GNNs from model extraction attacks
Researchers have developed GraphRP, a novel defense framework designed to protect Graph Neural Networks (GNNs) from model extraction attacks. Existing methods often fail due to an "Euclidean bias" that doesn't account f…
-
New theory quantifies information loss in quantum circuits
Researchers have developed a theoretical framework to understand how information is lost in parameterized quantum circuits when using a fixed computational-basis measurement. They analyzed the quantum Fisher information…
-
Catastrophic forgetting in LLMs: How fine-tuning erodes capabilities
Fine-tuning large language models can lead to catastrophic forgetting, where a model loses previously acquired capabilities when optimized for a new objective. This phenomenon, rooted in gradient descent, causes the mod…
-
GFlowNet training methods explored via policy gradients and information geometry · 2 sources tracked
Two new research papers explore advanced training methods for Generative Flow Networks (GFlowNets). The first paper introduces a policy-gradient-based framework that bridges GFlowNet's flow balance with reinforcement le…
-
CalTwin improves medical world models with Fisher-Information regularization
Researchers have developed CalTwin, a novel regularization technique designed to improve the reliability of medical world models. This method addresses two key issues: covariate shift, which arises from data fragmentati…
-
New Bayesian Optimization Method RAMBO Tackles Multi-Regime Search Spaces
Researchers have developed a new Bayesian Optimization (BO) method called RAMBO, designed to handle multi-regime search spaces more effectively than standard BO. RAMBO utilizes a Dirichlet Process Mixture of Gaussian Pr…
-
New Fisher Widths Analyze Statistical Manifold Complexity
This paper introduces two new functionals, the primal Fisher width and the inverse-Fisher width, to analyze Gaussian-width complexity on statistical manifolds. These widths offer complementary insights into local parame…
-
New math paper details optimization on Wasserstein space
Researchers have developed new methods for optimizing functionals on Wasserstein spaces, a mathematical concept crucial for understanding probability distributions. The work establishes linear convergence for proximal d…
-
CoCurve method prunes LLMs by analyzing cross-module dependencies
Researchers have developed CoCurve, a novel training-free method for structured pruning of large language models (LLMs). Unlike previous methods that assess computational units independently, CoCurve considers the inter…
-
New GFWSVD method offers improved LLM compression by analyzing parameter correlations
Researchers have developed Generalized Fisher-Weighted SVD (GFWSVD), a novel technique for compressing large language models (LLMs). This method improves upon existing compression methods by accounting for both diagonal…
-
New Geometric Observability Index Enhances SE(3) Pose Estimation
Researchers have introduced the Geometric Observability Index (GOI), a novel metric for assessing the sensitivity of pose estimation in SE(3) environments. This index quantifies the influence of individual measurements …
-
New CuBAS framework uses information geometry for adaptive data sampling
Researchers have developed CuBAS (Curvature-Based Adaptive Sampling), a novel framework for selecting informative data points for supervised classification tasks. This method leverages information geometry, viewing a la…
-
OrthoTryOn framework enhances unified fashion generation by resolving task conflicts
Researchers have developed OrthoTryOn, a novel framework designed to improve unified fashion generation models. This approach tackles the issue of negative transfer and gradient conflict that arises when multiple distin…