A naukri.com initiative
ML News
Arxiv
4d
178
Image Credit: Arxiv
Influential Bandits: Pulling an Arm May Change the Environment
- A new paper published on arXiv proposes the influential bandit problem, a multi-armed bandit formulation that considers interdependencies and non-stationary environments.
- The problem models arm interactions through an unknown interaction matrix that governs the dynamics of arm losses.
- The paper establishes regret lower bounds for standard bandit algorithms and introduces a new algorithm based on a lower confidence bound (LCB) estimator.
- Empirical evaluations demonstrate the presence of inter-arm influence and confirm the superior performance of the proposed method compared to conventional bandit algorithms.
Read Full Article
10 Likes
Arxiv
4d
332
Image Credit: Arxiv
DrivAer Transformer: A high-precision and fast prediction method for vehicle aerodynamic drag coefficient based on the DrivAerNet++ dataset
- A new study proposes the DrivAer Transformer (DAT), a point cloud learning framework for evaluating vehicle aerodynamic performance.
- The DAT structure uses the DrivAerNet++ dataset, containing high-fidelity CFD data of 3D vehicle shapes.
- DAT enables accurate estimation of air drag directly from 3D meshes, avoiding limitations of traditional methods.
- The framework is expected to accelerate the vehicle design process and improve development efficiency.
Read Full Article
20 Likes
Arxiv
4d
304
Image Credit: Arxiv
Millions of States: Designing a Scalable MoE Architecture with RWKV-7 Meta-learner
- State-based sequence models like RWKV-7 offer a compelling alternative to Transformer architectures.
- RWKV-7 lacks mechanisms for token-parameter interactions and native scalability.
- A novel extension to RWKV-7 called Meta-State is proposed, which replaces attention mechanisms with a fully state-driven approach.
- Meta-State supports progressive model scaling and offers a flexible framework for efficient and adaptable sequence modeling.
Read Full Article
18 Likes
Arxiv
4d
276
Image Credit: Arxiv
Enabling Automatic Differentiation with Mollified Graph Neural Operators
- Physics-informed neural operators offer a powerful framework for learning solution operators of partial differential equations (PDEs) by combining data and physics losses.
- The mollified graph neural operator (mGNO) is introduced as the first method to leverage automatic differentiation and compute exact gradients on arbitrary geometries.
- mGNO enables efficient training on irregular grids and varying geometries, while allowing seamless evaluation of physics losses at randomly sampled points for improved generalization.
- mGNOs demonstrate superior performance compared to finite differences and machine learning baselines when solving PDEs on regular and unstructured grids.
Read Full Article
16 Likes
Arxiv
4d
154
Image Credit: Arxiv
SortBench: Benchmarking LLMs based on their ability to sort lists
- Sorting is a challenging task for Large Language Models (LLMs) due to weaknesses in faithfully representing input data, logical comparisons, and differentiating between syntax and semantics.
- A new benchmark called SortBench for LLMs has been introduced, offering various difficulty levels and easy scalability.
- Tests conducted on seven state-of-the-art LLMs, including test-time reasoning models, revealed that even highly capable models like o3-mini can struggle with sorting tasks that involve mixing syntax and semantics.
- The models also face difficulties in preserving the faithfulness to input for long lists, often dropping or adding items. Test-time reasoning tends to overthink problems, leading to performance degradation.
Read Full Article
9 Likes
Arxiv
4d
284
Image Credit: Arxiv
Academic Network Representation via Prediction-Sampling Incorporated Tensor Factorization
- Accurate representation to an academic network is of great significance to academic relationship mining like predicting scientific impact.
- The paper proposes a Prediction-sampling-based Latent Factorization of Tensors (PLFT) model to address the issue of high-dimensional and incomplete academic networks.
- The PLFT model includes a cascade LFT architecture to enhance model representation learning ability and a predicting-sampling strategy to more accurately learn the network representation.
- Experimental results show that the PLFT model outperforms existing models in predicting unexplored relationships in academic networks.
Read Full Article
17 Likes
Arxiv
4d
304
Image Credit: Arxiv
Towards generalizable single-cell perturbation modeling via the Conditional Monge Gap
- Learning the response of single-cells to various treatments offers great potential to enable targeted therapies.
- Neural optimal transport (OT) has emerged as a methodological framework for analyzing unpaired single-cell data induced by cell destruction during data acquisition.
- The Conditional Monge Gap is proposed as a method that learns OT maps conditionally on arbitrary covariates, such as time, drug treatment, drug dosage, or cell type.
- The conditional models show promising generalization performance to unseen treatments, outperforming other models in capturing heterogeneity in the perturbed population.
Read Full Article
18 Likes
Arxiv
4d
277
Image Credit: Arxiv
An Adaptive Clustering Scheme for Client Selections in Communication-Efficient Federated Learning
- Federated learning is a decentralized learning architecture that consumes a lot of network transmission resources.
- A new adaptive clustering scheme is proposed to reduce communication costs in federated learning.
- The scheme dynamically adjusts the number of clusters to find the most ideal grouping results.
- Experimental results show a reduction of communication and transmission costs by almost 50% without affecting model accuracy.
Read Full Article
16 Likes
Arxiv
4d
385
Image Credit: Arxiv
DRIP: DRop unImportant data Points -- Enhancing Machine Learning Efficiency with Grad-CAM-Based Real-Time Data Prioritization for On-Device Training
- Effective selection methods for model training can reduce labeling effort, optimize on-device training, and enhance model performance.
- A novel algorithm using Grad-CAM is introduced for online decision-making on data point retention or discarding.
- The algorithm computes a unique DRIP Score to quantify the importance of each data point.
- Experimental evaluations show that the approach achieves storage savings of up to 39% while maintaining or surpassing model accuracy.
Read Full Article
23 Likes
Arxiv
4d
121
Image Credit: Arxiv
Proofs as Explanations: Short Certificates for Reliable Predictions
- This research explores a model for explainable AI that provides proof as explanations for reliable predictions.
- The model defines an explanation as a subset of the training data that can serve as a proof of a prediction's correctness.
- The research presents the concept of the robust hollow star number to determine the worst-case size of the smallest certificate achievable.
- The study also analyzes worst-case distributional bounds and distribution-dependent bounds for certificate size.
Read Full Article
7 Likes
Arxiv
4d
121
Image Credit: Arxiv
Scaling Up On-Device LLMs via Active-Weight Swapping Between DRAM and Flash
- Large language models (LLMs) are being deployed on mobile devices, but limited DRAM capacity constrains the model size.
- ActiveFlow is introduced as an LLM inference framework that enables adaptive DRAM usage for modern LLMs.
- ActiveFlow utilizes novel techniques such as cross-layer active weights preloading and sparsity-aware self-distillation.
- The framework achieves the performance-cost Pareto frontier compared to existing optimization methods.
Read Full Article
7 Likes
Arxiv
4d
240
Image Credit: Arxiv
PCA-RAG: Principal Component Analysis for Efficient Retrieval-Augmented Generation
- Retrieval-Augmented Generation (RAG) is a powerful paradigm for grounding large language models in external knowledge sources.
- This paper explores the use of Principal Component Analysis (PCA) to reduce the dimensionality of language model embeddings, addressing scalability challenges in processing large financial text corpora.
- By reducing vectors from 3,072 to 110 dimensions, significant speedup in retrieval operations and reduction in index size are achieved, with moderate declines in correlation metrics.
- The study highlights the practicality of leveraging classical dimensionality reduction techniques to optimize RAG architectures for knowledge-intensive applications in finance and trading.
Read Full Article
14 Likes
Arxiv
4d
272
Image Credit: Arxiv
Graph Reduction with Unsupervised Learning in Column Generation: A Routing Application
- Column Generation (CG) is a popular method for enhancing computational efficiency in large scale Combinatorial Optimization problems.
- A new approach combines CG with Graph Neural Network (GNN) and unsupervised learning to reduce the size of the Elementary Shortest Path Problem with Resource Constraints (ESPPRC).
- The reduced problem is then solved using local search techniques, resulting in significant improvements in convergence compared to previous reduction techniques.
- The method has shown over 9% improvement in objective values for larger instances of Capacitated Vehicle Routing Problems with Time Windows.
Read Full Article
16 Likes
Arxiv
4d
300
Image Credit: Arxiv
BOISHOMMO: Holistic Approach for Bangla Hate Speech
- A multi-label Bangla hate speech dataset named BOISHOMMO has been compiled and evaluated.
- BOISHOMMO includes categories of hate speech across various dimensions such as race, gender, religion, and politics.
- The dataset consists of over two thousand annotated examples, providing a nuanced understanding of hate speech in Bangla.
- The dataset aims to improve hate speech detection and analysis studies for low-resource languages.
Read Full Article
18 Likes
Arxiv
4d
8
Image Credit: Arxiv
Constrained Machine Learning Through Hyperspherical Representation
- The problem of ensuring constraints satisfaction on the output of machine learning models is critical for many applications, especially in safety-critical domains.
- A novel method called Hypersherical Constrained Representation is proposed to enforce constraints in the output space for convex and bounded feasibility regions.
- The method operates on a different representation system, where Euclidean coordinates are converted into hyperspherical coordinates relative to the constrained region, thereby ensuring only feasible points are represented.
- Experiments on synthetic and real-world datasets show that the proposed method achieves comparable predictive performance, guarantees 100% constraint satisfaction, and has minimal computational cost at inference time.
Read Full Article
Like
For uninterrupted reading, download the app