A naukri.com initiative
ML News
Arxiv
1d
18
Image Credit: Arxiv
Effortless, Simulation-Efficient Bayesian Inference using Tabular Foundation Models
- Simulation-based inference (SBI) uses neural networks to rapidly infer posterior distributions for observed data.
- Tabular foundation models, such as TabPFN, can be used as pre-trained autoregressive conditional density estimators for SBI.
- Neural Posterior Estimation with Prior-data Fitted Networks (NPE-PF) is competitive in terms of accuracy and simulation efficiency.
- NPE-PF eliminates the need for inference network selection, training, and hyperparameter tuning.
Read Full Article
Like
Arxiv
1d
308
Image Credit: Arxiv
Federated Learning: A Survey on Privacy-Preserving Collaborative Intelligence
- Federated Learning (FL) is a transformative paradigm in the field of distributed machine learning.
- FL enables multiple clients to collaboratively train a shared global model without centralizing sensitive data.
- This survey provides an overview of FL, covering architecture, communication protocol, challenges, and real-world applications.
- It also highlights open research problems and future directions in developing scalable and trustworthy FL systems.
Read Full Article
18 Likes
Arxiv
1d
330
Image Credit: Arxiv
Fault Diagnosis in New Wind Turbines using Knowledge from Existing Turbines by Generative Domain Adaptation
- Intelligent condition monitoring of wind turbines is essential for reducing downtimes. Machine learning models trained on wind turbine operation data are commonly used to detect anomalies and operation faults.
- A novel generative deep learning approach is presented to make SCADA samples from wind turbines with limited training data resemble those with representative training data.
- The proposed technique improves fault diagnosis in wind turbines with scarce data, achieving similar anomaly scores to models trained with abundant data.
- This research direction provides a promising solution for improving anomaly detection and fault diagnosis in wind farms with limited training data.
Read Full Article
19 Likes
Arxiv
1d
63
Image Credit: Arxiv
Early Detection of Multidrug Resistance Using Multivariate Time Series Analysis and Interpretable Patient-Similarity Representations
- Researchers propose an interpretable Machine Learning (ML) framework for Multidrug Resistance (MDR) prediction.
- The framework models patients as Multivariate Time Series (MTS) and uses various similarity measures to quantify patient interactions.
- It achieves an AUC of 81% and outperforms baseline ML and deep learning models in MDR prediction.
- The approach identifies key risk factors and reveals clinically relevant clusters, supporting early detection and patient stratification.
Read Full Article
3 Likes
Arxiv
1d
157
Image Credit: Arxiv
Conformal Segmentation in Industrial Surface Defect Detection with Statistical Guarantees
- Traditional surface defect detection methods in industrial settings rely on manual inspection, which is inefficient and costly.
- Automated defect detection approaches using Convolutional Neural Networks have limitations due to data annotation uncertainties and overfitting issues.
- A statistically rigorous threshold is derived to identify high-probability defective pixels in test images, ensuring reliable detection.
- The study demonstrates control over the expected test set error rate, validating the method's adaptability and effectiveness.
Read Full Article
9 Likes
Arxiv
1d
165
Image Credit: Arxiv
Towards Robust LLMs: an Adversarial Robustness Measurement Framework
- The rise of Large Language Models (LLMs) has revolutionized artificial intelligence, but they are vulnerable to adversarial perturbations.
- The Robustness Measurement and Assessment (RoMA) framework is adapted to quantify LLM resilience against adversarial inputs.
- RoMA's estimates demonstrate accuracy with minimal error margins and computational efficiency.
- LLM robustness varies between models, categories within the same task, and types of perturbations, highlighting the need for task-specific evaluations.
Read Full Article
9 Likes
Arxiv
1d
47
Image Credit: Arxiv
MSGCN: Multiplex Spatial Graph Convolution Network for Interlayer Link Weight Prediction
- Graph Neural Networks (GNNs) have been widely used for various learning tasks, including link prediction.
- MSGCN (Multiplex Spatial Graph Convolution Network) is a new method that predicts interlayer link weights in multilayer networks.
- MSGCN spatially embeds information across multiple layers and captures the geometric structure of nodes.
- Extensive experiments demonstrate the robust and accurate link weight prediction performance of the MSGCN model.
Read Full Article
2 Likes
Arxiv
1d
135
Image Credit: Arxiv
Disaggregated Deep Learning via In-Physics Computing at Radio Frequency
- WISE is a novel computing architecture for wireless edge networks designed to overcome energy constraints in deep learning inference.
- It achieves this through two key innovations: disaggregated model access via wireless broadcasting and in-physics computation of complex-valued matrix-vector multiplications directly at radio frequency.
- WISE achieves 95.7% image classification accuracy with ultra-low operation power of 6.0 fJ/MAC per client, resulting in a computation efficiency of 165.8 TOPS/W.
- This approach enables energy-efficient deep learning inference on wirelessly connected edge devices, achieving over two orders of magnitude improvement compared to traditional digital computing.
Read Full Article
8 Likes
Arxiv
1d
139
Image Credit: Arxiv
Flexibility of German gas-fired generation: evidence from clustering empirical operation
- A study conducted on German national gas generation units revealed variations in their flexibility.
- Over 60% of the gas generation units were classified into clusters based on their empirical flexibility.
- Two clusters consisted of peaker units while the other two clusters comprised non-peaker units.
- Non-peaker units, which are less flexible, account for more than 83% of the sample must-run generation.
Read Full Article
8 Likes
Arxiv
1d
266
Image Credit: Arxiv
Bidirectional Mamba for Single-Cell Data: Efficient Context Learning with Biological Fidelity
- GeneMamba is introduced as a scalable and efficient foundation model for single-cell transcriptomics built on state space modeling.
- It captures bidirectional gene context with linear-time complexity, offering substantial computational gains over transformer baselines.
- GeneMamba is pretrained on nearly 30 million cells and incorporates biologically informed objectives, including pathway-aware contrastive loss and rank-based gene encoding.
- Evaluation of GeneMamba showcases its strong performance, interpretability, and robustness, making it a practical and powerful alternative to transformer-based methods for single-cell data analysis.
Read Full Article
16 Likes
Arxiv
1d
18
Image Credit: Arxiv
Engineering the Law-Machine Learning Translation Problem: Developing Legally Aligned Models
- Organizations developing machine learning-based (ML) technologies face the challenge of achieving high predictive performance while respecting the law.
- ML model behavior cannot be directly operationalized in source code to meet legal obligations.
- A five-stage interdisciplinary framework is introduced to integrate legal and ML-technical analysis during ML model development.
- The framework helps design legally aligned ML models and identify high-performing models that are legally justifiable.
Read Full Article
1 Like
Arxiv
1d
218
Image Credit: Arxiv
A Systematic Approach to Design Real-World Human-in-the-Loop Deep Reinforcement Learning: Salient Features, Challenges and Trade-offs
- A novel multi-layered hierarchical HITL DRL algorithm is introduced for decision-making problems.
- The algorithm combines self learning, imitation learning, and transfer learning.
- Human inputs in the form of reward, action, and demonstration are considered.
- Multiple challenges, trade-offs, and advantages of HITL are discussed in solving complex problems.
Read Full Article
13 Likes
Arxiv
1d
338
Image Credit: Arxiv
Neural Theorem Proving: Generating and Structuring Proofs for Formal Verification
- Formally verifying properties of software code, especially with the emergence of LLM-generated code, is a highly desirable task.
- This work introduces a framework that generates whole proofs in a formal language to be used in systems utilizing built-in tactics and automated theorem provers.
- The framework includes components for generating natural language statements, an LLM that generates formal proofs, and a module employing heuristics for building the final proof.
- The framework is validated using benchmark tests and the Isabelle proof assistant, and a dataset is curated for future training tasks.
Read Full Article
20 Likes
Arxiv
1d
56
Image Credit: Arxiv
Neural Contraction Metrics with Formal Guarantees for Discrete-Time Nonlinear Dynamical Systems
- This paper introduces the approach of learning verifiable contraction metrics for discrete-time nonlinear dynamical systems.
- The paper addresses the challenge of identifying contraction metrics for complex nonlinear systems using neural networks.
- A new sufficient condition is established for formal neural contraction metrics, assuming continuity of the dynamics.
- The paper validates the approach through successful synthesis and verification of neural contraction metrics for various nonlinear examples.
Read Full Article
3 Likes
Arxiv
1d
311
Image Credit: Arxiv
PACE: A Framework for Learning and Control in Linear Incomplete-Information Differential Games
- Researchers propose the Peer-Aware Cost Estimation (PACE) framework for learning the cost parameters of another agent in a linear quadratic differential game with incomplete information.
- PACE treats the other agent as a learning agent rather than a stationary optimal agent and models their learning dynamics to infer their cost function parameters.
- The PACE framework enables agents to adapt their control policies based on real-time inference of each other's objective functions, using only previous state observations.
- Numerical studies show that modeling the learning dynamics of the other agent improves stability and convergence speed compared to approaches assuming complete information.
Read Full Article
18 Likes
For uninterrupted reading, download the app