A naukri.com initiative
ML News
Arxiv
2h
282
Image Credit: Arxiv
A Privacy-Preserving Federated Learning Framework for Generalizable CBCT to Synthetic CT Translation in Head and Neck
- Cone-beam computed tomography (CBCT) for image-guided radiotherapy lacks soft-tissue contrast and accuracy in dose calculation, leading to interest in Synthetic CT (sCT) generation using deep learning methods.
- A privacy-preserving federated learning (FL) framework was proposed for CBCT-to-sCT translation in head and neck regions, showcasing cross-silo horizontal FL approach.
- A generative adversarial network was collaboratively trained on data from three European medical centers, demonstrating good generalization across centers with promising results on an external validation dataset.
- The federated model achieved comparable performance on the validation dataset without additional training, highlighting the potential of FL for CBCT-to-sCT synthesis across institutions while preserving data privacy.
Read Full Article
17 Likes
Arxiv
2h
94
Image Credit: Arxiv
sparseGeoHOPCA: A Geometric Solution to Sparse Higher-Order PCA Without Covariance Estimation
- sparseGeoHOPCA is a new framework for sparse higher-order principal component analysis (SHOPCA) that utilizes a geometric approach for tensor decomposition.
- The method converts the original nonconvex sparse objective into a manageable geometric form by restructuring subproblems as structured binary linear optimization problems.
- sparseGeoHOPCA eliminates the need for covariance estimation and iterative deflation, leading to improved computational efficiency and interpretability, especially in high-dimensional and unbalanced data scenarios.
- The algorithm achieves a total computational complexity that scales linearly with tensor size and has been shown to accurately recover sparse supports, maintain classification performance under compression, and offer high-quality image reconstruction on ImageNet.
Read Full Article
5 Likes
Arxiv
2h
169
Image Credit: Arxiv
ConfPO: Exploiting Policy Model Confidence for Critical Token Selection in Large Language Model Preference Optimization
- ConfPO is a method for preference learning in Large Language Models (LLMs) that optimizes preference-critical tokens based on the training policy's confidence.
- ConfPO focuses on optimizing the most impactful tokens, improving alignment quality and mitigating overoptimization compared to prior Direct Alignment Algorithms (DAAs).
- ConfPO does not require auxiliary models or additional compute, making it a simple, lightweight, and model-free approach.
- Experimental results show that ConfPO consistently outperforms uniform DAAs on various LLMs, achieving better alignment with zero extra computational costs.
Read Full Article
10 Likes
Arxiv
2h
58
Image Credit: Arxiv
Stop Misusing t-SNE and UMAP for Visual Analytics
- Misuses of t-SNE and UMAP in visual analytics have become increasingly common.
- Practitioners often use t-SNE and UMAP projections to investigate inter-cluster relationships even though these projections may not reflect true distances between clusters accurately.
- The misuse of t-SNE and UMAP is attributed to limited discourse on their appropriate use in visual analytics.
- A literature review of 114 papers and an interview study were conducted to verify the prevalence of misuse and understand practitioners' motivations for using these techniques.
Read Full Article
3 Likes
Arxiv
2h
286
Image Credit: Arxiv
Flexible and Efficient Drift Detection without Labels
- Machine learning models are increasingly used for automated decision-making, requiring early detection of concept drift for optimal performance.
- Current research on concept drift mainly focuses on supervised tasks with immediate access to true labels, posing challenges for large datasets without instant labels.
- A new algorithm utilizing statistical process control in a label-less setting is proposed for efficient concept drift detection with improved statistical power.
- Introduction of a novel drift detection framework enhances the algorithm's performance in detecting drift without labels, as demonstrated through numerical simulations.
Read Full Article
17 Likes
Arxiv
2h
188
Image Credit: Arxiv
Bridging RDF Knowledge Graphs with Graph Neural Networks for Semantically-Rich Recommender Systems
- Graph Neural Networks (GNNs) have greatly advanced recommender systems, but have not fully utilized the semantic information in knowledge graphs (KGs) like RDF.
- A new approach integrates RDF KGs with GNNs by leveraging topological and content information from RDF object and datatype properties.
- The study evaluates various GNNs, analyzing how semantic feature initializations and graph structure heterogeneity affect their performance in recommendation tasks.
- Experiments on multi-million-node RDF graphs show that leveraging RDF KGs' semantic richness significantly enhances recommender systems, paving the way for GNN-based systems in Linked Open Data cloud.
Read Full Article
11 Likes
Arxiv
2h
178
Image Credit: Arxiv
Towards Secure and Private Language Models for Nuclear Power Plants
- Introduction of a domain-specific Large Language Model for nuclear applications, based on the Essential CANDU textbook, to protect sensitive data in nuclear operations.
- Model uses a compact Transformer-based architecture, trained on a single GPU, showcasing understanding of specialized nuclear vocabulary with some limitations in syntactic coherence.
- Focus on in-house LLM solutions for cybersecurity and data confidentiality standards, highlighting early successes in text generation and the need for improvements in dataset coverage and preprocessing.
- Future directions include expanding the dataset to cover diverse nuclear subtopics, improving tokenization, and evaluating readiness of the model for practical use in the nuclear domain.
Read Full Article
10 Likes
Arxiv
2h
169
Image Credit: Arxiv
Superposed Parameterised Quantum Circuits
- Quantum machine learning has limitations due to linear unitary operations and shared trainable parameters across outputs.
- Superposed parameterised quantum circuits are introduced to overcome these limitations by embedding an exponential number of parameterised sub-models in a single circuit.
- This new architecture induces polynomial activation functions through amplitude transformations and post-selection, allowing for training multiple parameter sets in parallel.
- Numerical experiments show significant advantages of superposed parameterised quantum circuits in reducing errors and improving accuracy in tasks like regression and classification.
Read Full Article
10 Likes
Arxiv
2h
74
Image Credit: Arxiv
Enhancing Accuracy and Maintainability in Nuclear Plant Data Retrieval: A Function-Calling LLM Approach Over NL-to-SQL
- Retrieving operational data from nuclear power plants requires exceptional accuracy and transparency.
- Traditional NL-to-SQL approaches for querying such data pose risks due to difficulties in validation and complex database structures.
- A new approach leveraging function-calling large language models (LLMs) is proposed to enhance accuracy and maintainability.
- The hybrid approach defines pre-approved functions for querying, ensuring SQL queries are validated and optimized by experts before deployment.
Read Full Article
4 Likes
Arxiv
2h
315
Image Credit: Arxiv
EDINET-Bench: Evaluating LLMs on Complex Financial Tasks using Japanese Financial Statements
- A new open-source Japanese financial benchmark called EDINET-Bench has been introduced to evaluate the performance of large language models (LLMs) on complex financial tasks like accounting fraud detection, earnings forecasting, and industry prediction.
- EDINET-Bench is constructed by gathering annual reports from the past 10 years from Japan's Electronic Disclosure for Investors' NETwork (EDINET) and automatically assigning labels for evaluation tasks.
- Experiments indicate that even the best LLMs struggle in performing better than logistic regression in binary classification for fraud detection and earnings forecasting using the EDINET-Bench dataset.
- The study emphasizes the challenges of applying LLMs to practical financial applications and suggests the necessity for domain-specific adaptation. The dataset, benchmark construction code, and evaluation code are made publicly available for further research.
Read Full Article
18 Likes
Arxiv
2h
217
Image Credit: Arxiv
Paths to Causality: Finding Informative Subgraphs Within Knowledge Graphs for Knowledge-Based Causal Discovery
- Inferring causal relationships between variables is essential for understanding multivariate interactions in complex systems.
- Knowledge-based causal discovery relies on reasoning over metadata of variables, offering an alternative to traditional observational data methods.
- A novel approach integrating Knowledge Graphs with Large Language Models improves knowledge-based causal discovery by identifying informative subgraphs and refining their selection.
- Extensive experiments on biomedical and open-domain datasets show that this method outperforms baselines in inferring causal relationships, offering a significant improvement in F1 scores.
Read Full Article
13 Likes
Arxiv
2h
110
Image Credit: Arxiv
syren-baryon: Analytic emulators for the impact of baryons on the matter power spectrum
- Baryonic physics significantly influences the distribution of matter in the Universe on cosmological scales, making it a crucial factor in cosmological surveys.
- Researchers have developed symbolic parametrizations to model the impact of baryonic physics on the matter power spectrum for various models, considering parameters such as wavenumber, redshift, cosmology, and baryonic feedback control.
- The study utilizes symbolic regression to create analytic approximations for the ratio of the matter power spectrum with baryons to that without, considering different sub-grid prescriptions of baryonic physics from hydrodynamical simulations.
- Their analytic expressions allow for a direct understanding of how changing cosmological and feedback parameters affects the models, enabling distinctions between different baryonic physics implementations when analyzing real data.
Read Full Article
6 Likes
Arxiv
2h
123
Image Credit: Arxiv
On The Impact of Merge Request Deviations on Code Review Practices
- Code review is crucial in software engineering for maintaining quality and collaboration.
- Many industrial Merge Request (MR) workflows deviate from standard review processes, with a significant percentage serving non-review purposes.
- Identified seven categories of deviations in MRs, occurring in 37.02% of cases, and proposed a few-shot learning detection method with 91% accuracy.
- Excluding deviations improves ML models predicting review completion time, enhancing performance and shifting feature importance significantly.
Read Full Article
7 Likes
Arxiv
2h
117
Image Credit: Arxiv
StreamSplat: Towards Online Dynamic 3D Reconstruction from Uncalibrated Video Streams
- Real-time reconstruction of dynamic 3D scenes from uncalibrated video streams is important for real-world applications.
- StreamSplat is introduced as a fully feed-forward framework for transforming uncalibrated video streams into dynamic 3D Gaussian Splatting representations online.
- Key technical innovations include a probabilistic sampling mechanism in the static encoder and a bidirectional deformation field in the dynamic decoder for robust and efficient dynamic modeling.
- StreamSplat has shown better performance in reconstruction quality and dynamic scene modeling compared to previous methods, and supports online reconstruction of long video streams.
Read Full Article
7 Likes
Arxiv
2h
230
Image Credit: Arxiv
AdversariaL attacK sAfety aLIgnment(ALKALI): Safeguarding LLMs through GRACE: Geometric Representation-Aware Contrastive Enhancement- Introducing Adversarial Vulnerability Quality Index (AVQI)
- Adversarial threats against LLMs are evolving faster than current defenses can adapt, showing a critical geometric blind spot in alignment.
- Introducing ALKALI, a benchmark with 9,000 prompts across various attack families to assess the vulnerability of 21 leading LLMs, highlighting high Attack Success Rates (ASRs).
- To address the vulnerability of latent camouflage, GRACE - Geometric Representation Aware Contrastive Enhancement is introduced, reducing ASR by up to 39% through preference learning and latent space regularization.
- AVQI, a geometry-aware metric, is introduced to quantify latent alignment failure by measuring cluster separation and compactness, providing insights into how models encode safety internally.
Read Full Article
13 Likes
For uninterrupted reading, download the app