singular value decomposition
PulseAugur coverage of singular value decomposition — every cluster mentioning singular value decomposition across labs, papers, and developer communities, ranked by signal.
19 day(s) with sentiment data
-
New defense system ORAN-DEFEND targets backdoor attacks in Open RAN
Researchers have developed ORAN-DEFEND, a new system designed to protect Open Radio Access Networks (O-RAN) from backdoor attacks embedded in third-party deep reinforcement learning (DRL) xApps. This defense mechanism o…
-
New MESH-FL framework boosts federated learning compression on edge devices
Researchers have developed MESH-FL, a novel framework for federated learning on edge devices that utilizes entropy-guided compression for multimodal models. This approach adaptively allocates compression ranks across di…
-
New framework empirically compresses deep neural networks via state analysis
Researchers have developed a novel method for compressing deep neural networks by analyzing the controllability and observability of their internal states. This framework treats trained networks as dynamical systems, us…
-
LoCA method adapts vision foundation models efficiently for convolutional layers
Researchers have introduced LoCA (Low-Rank Convolutional Adaptation), a novel method for efficiently fine-tuning vision foundation models. Unlike existing LoRA techniques that are primarily designed for transformer arch…
-
New research explores advanced LLM compression techniques · 2 sources tracked
Two new research papers propose advanced methods for compressing large language models (LLMs) to reduce their size and computational requirements. The first paper introduces Leech Lattice Vector Quantization (LLVQ), whi…
-
New 'in-span learning' method adapts AI models using their own predictions
A new research paper introduces "in-span learning," a method for adapting reduced-order models using their own predictions. This technique enhances the model's ability to absorb future corrections by reweighting and rea…
-
New CORA method cuts LLM fine-tuning parameters by 4x
Researchers have introduced CORA (Coherent Orthogonal Rotation Adaptation), a novel parameter-efficient fine-tuning method for large language models. CORA leverages singular value decomposition (SVD) to preserve the geo…
-
New TAO method improves robot visual place recognition accuracy
Researchers have developed Trajectory-Anchor Optimization (TAO), a new method to improve thermal visual place recognition (TIR-VPR) in robots. Existing foundation model-based TIR-VPR systems can be overconfident, falsel…
-
New framework reveals how Vision Transformers encode geometry
Researchers have developed a new framework to analyze how self-supervised Vision Transformers (ViTs) encode geometric information. By using Singular Value Decomposition (SVD) to examine the weights of linear probes, the…
-
CLEAR-MoE converts frozen Vision Transformers to sparse MoE models
Researchers have developed CLEAR-MoE, a novel post-training method to transform frozen Vision Transformers (ViTs) into sparse Mixture-of-Experts (MoE) models without altering the original backbone weights. This techniqu…
-
New Math Model for Neural Network Initialization Spectra
Researchers have developed a new mathematical framework to analyze the singular value spectrum of products of non-square random matrices. This framework is applicable to understanding the feature covariance eigenvalues …
-
New method reframes continual learning as retrieval-based memory management
Researchers have introduced Neural Subspace Reallocation (NSR), a novel approach to continual learning that frames the process as memory management within parameter subspaces. NSR treats Low-Rank Adaptation (LoRA) modul…
-
New S-GAI framework embeds dataset geometry into MLP weights
Researchers have developed S-GAI, a novel initialization framework for sigmoidal MLPs that embeds dataset geometry directly into network weights. This method uses singular value decomposition (SVD) to estimate class-wis…
-
New ReTeX framework recovers task expert performance from merged AI models
Researchers have developed a new framework called ReTeX to address parameter interference in multi-task model merging. This method models interference as additive offsets and predicts these offsets to recover individual…
-
New hybrid method enhances privacy in semantic search
Researchers have developed a novel approach to privacy-aware semantic search that balances data protection with search performance. This method uses Singular Value Decomposition (SVD) to truncate document embeddings int…
-
New hybrid method enhances privacy in semantic search systems
Researchers have developed a novel approach to enhance privacy in semantic search systems, which are powered by dense embeddings. The proposed method addresses the risk of embedding-inversion attacks that can reconstruc…
-
New SVD framework tackles shift-variant image blur
Researchers have developed a new framework using singular value decomposition (SVD) to restore images affected by shift-variant motion blur. This method addresses the challenge of varying degradation across an image, wh…
-
New SVD-Surgeon method optimizes LLM compression without retraining
Researchers have developed SVD-Surgeon, a novel training-free method for compressing large language models (LLMs) using singular value decomposition (SVD). This technique optimizes the singular values directly, offering…
-
New method merges AI models without training for better performance
Researchers have developed a novel training-free method for merging multiple task-specific AI models into a single, more efficient multi-task model. This new approach, called SiM, uses singular value decomposition to ap…
-
New ORBIT method enables multi-attribute steering in language models
Researchers have developed ORBIT, a novel training-free method for simultaneously steering multiple behavioral attributes in language models. Unlike previous methods that struggle with combining attributes or require re…