Jacobian matrix
PulseAugur coverage of Jacobian matrix — every cluster mentioning Jacobian matrix across labs, papers, and developer communities, ranked by signal.
6 day(s) with sentiment data
-
New Target-Response Jacobian metric outperforms influence measures in AI model analysis
Researchers have introduced the target-response Jacobian ($J$), a new metric for understanding how model predictions change in response to alterations in individual training targets. This metric complements existing tra…
-
New Rényi-based metrics enhance MLP interpretability
Researchers have introduced a new family of metrics, termed \"alpha-RC metrics,\" designed to enhance the interpretability of multilayer perceptrons (MLPs). These metrics are based on Rényi divergence and aim to quantif…
-
New LeMA framework enables scalable Levenberg-Marquardt training for large deep learning models
Researchers have developed LeMA, a distributed training framework for deep learning models that makes the Levenberg-Marquardt (LMA) algorithm practical for large-scale neural networks. LeMA addresses the prohibitive mem…
-
New JAX toolkit enables efficient Automatic Sparse Differentiation
A new toolkit called asdex has been released for the JAX ecosystem, offering the first standalone implementation of Automatic Sparse Differentiation (ASD). This method is designed to efficiently compute Jacobian or Hess…
-
New research explores Gauss-Newton accuracy and indefinite Hessians in low-cost sets
This paper delves into the accuracy of the Gauss-Newton curvature within ridge-regularized nonlinear least squares. It establishes that under specific conditions of local regularity and level-set curvature persistence, …
-
Gauss-Newton Accuracy and Indefinite Hessians Explored in New Paper
This paper delves into the accuracy of the Gauss-Newton curvature within ridge-regularized nonlinear least squares. It establishes that under specific conditions of local regularity and persistent level-set curvature, t…
-
New analysis details differential privacy for gradient descent on perturbed objectives
Researchers have developed a method to analyze the differential privacy of gradient descent when applied to perturbed objectives. This approach involves adding a random linear term to the objective function before optim…
-
New algorithm simplifies federated bilevel optimization using first-order gradients
Researchers have developed a new algorithm for federated stochastic bilevel optimization that avoids the need for computationally expensive second-order Hessian and Jacobian matrices. This novel approach, termed federat…
-
Robotics research introduces outcome-based representation learning
Researchers have developed a novel method for learning manipulation-sufficient representations in robotics, focusing on action outcomes rather than dense geometric states. This approach utilizes an action-conditioned ou…
-
New Jacobian Lens Tool Tests if AI Models Use Internal Signals
Researchers have developed a new interpretability tool called the Jacobian lens, designed to determine if a model's internal signals are actively used in its decision-making process. Unlike previous methods like the log…
-
New framework enhances power grid stability under uncertainty
Researchers have developed a new framework for ensuring the stability of AC optimal power flow systems under uncertainty. This method leverages a novel geometric approach, transforming stability requirements into convex…
-
New SC-NOs improve neural network accuracy for complex PDE modeling
Researchers have developed Sensitivity-Constrained Neural Operators (SC-NOs) to improve the reliability and data efficiency of neural networks used for modeling partial differential equations (PDEs). By incorporating Ja…
-
New theory explains Jacobian lens for language model interpretation
Researchers have developed a mathematical framework to better understand the Jacobian lens (J-lens), a method used to interpret representations within language models. The study provides a theoretical basis for the J-le…
-
New method enhances data utility under Local Differential Privacy
Researchers have developed a novel method to improve the utility of data collected under Local Differential Privacy (LDP). This approach selectively reduces noise in subspaces of the data that are most relevant to a spe…
-
New research shows normalization layers provide global context in sequence labelers
A new research paper challenges the conventional understanding of context in convolutional sequence labelers. The study demonstrates that sequence-pooled normalization layers can provide global context, bypassing the li…
-
Neural operators fail to reliably warm-start Newton solvers for PDEs
Researchers have identified a critical flaw in using neural operators to warm-start Newton solvers for nonlinear partial differential equations (PDEs). While neural operators can reduce test error, they may still produc…
-
FishBack method improves transformer activation steering using non-Euclidean geometry
Researchers have developed a new method called FishBack to improve activation steering in transformers, a technique for modifying language model behavior without updating parameters. Existing methods are often unstable …
-
New research offers theoretical convergence guarantees for neural network training
Two new research papers explore theoretical underpinnings of neural network training. The first paper establishes convergence guarantees for gradient descent in general feedforward neural networks by introducing a gener…
-
New research proposes unsupervised disentanglement via functional orthogonality
A new research paper proposes a novel approach to unsupervised disentangled representation learning by framing latent concepts as factors influencing observations through locally orthogonal directions. This method, form…
-
Gemma 2-2B research finds active feature planes have less holonomy
A new research paper published on arXiv investigates the concentration of holonomy within specific feature planes of the Gemma 2-2B model. The study preregistered its methodology and analysis rules before inspecting the…