A naukri.com initiative
ML News
Arxiv
4d
134
Image Credit: Arxiv
Rethinking Graph Structure Learning in the Era of LLMs
- Researchers are exploring the integration of language descriptions into graphs, known as text-attributed graphs (TAGs), to enhance model encoding capabilities.
- Graph structure learning (GSL) is a crucial technique for improving data utility, and it is highly relevant to efficient TAG learning.
- The challenge is to define a reasonable optimization objective for GSL in the era of large language models (LLMs) and design an efficient model architecture for LLM integration.
- The proposed Large Language and Tree Assistant (LLaTA) leverages tree-based LLM in-context learning to enhance the understanding of topology and text in order to generate improved graph structures.
Read Full Article
8 Likes
Arxiv
4d
138
Image Credit: Arxiv
Efficient Learning for Entropy-regularized Markov Decision Processes via Multilevel Monte Carlo
- Researchers propose efficient learning algorithms for entropy-regularized Markov Decision Processes (MDPs) with large or continuous state and action spaces.
- The algorithms integrate fixed-point iteration with multilevel Monte Carlo techniques and a stochastic approximation of the Bellman operator.
- Using a biased plain Monte Carlo estimate for the Bellman operator leads to quasi-polynomial sample complexity, while an unbiased randomized multilevel approximation achieves polynomial sample complexity in expectation.
- The proposed algorithms demonstrate performance guarantees independent of the dimensions or sizes of state and action spaces.
Read Full Article
8 Likes
Arxiv
4d
378
Image Credit: Arxiv
Feature-Enhanced Machine Learning for All-Cause Mortality Prediction in Healthcare Data
- Accurate patient mortality prediction enables effective risk stratification, leading to personalized treatment plans and improved patient outcomes.
- This study evaluates machine learning models for all-cause in-hospital mortality prediction using the MIMIC-III database, employing a comprehensive feature engineering approach.
- The Random Forest model achieved the highest performance with an AUC of 0.94, significantly outperforming other machine learning and deep learning approaches.
- The findings highlight the importance of careful feature engineering for accurate mortality prediction and propose future directions, including enhancing model robustness and tailoring prediction models for specific diseases.
Read Full Article
22 Likes
Arxiv
4d
390
Image Credit: Arxiv
Improving $(\alpha, f)$-Byzantine Resilience in Federated Learning via layerwise aggregation and cosine distance
- Researchers propose a new aggregation scheme, Layerwise Cosine Aggregation, to enhance the robustness of Federated Learning (FL) systems against Byzantine attacks.
- FL is a privacy-preserving approach for distributed machine learning, but it is vulnerable to malicious nodes contributing corrupted model updates.
- Layerwise Cosine Aggregation improves the performance of robust aggregation operators in high-dimensional parameter spaces, leading to up to a 16% increase in model accuracy.
- Theoretical analysis and empirical evaluation across various image classification datasets validate the superior robustness of Layerwise Cosine Aggregation.
Read Full Article
23 Likes
Arxiv
4d
97
Image Credit: Arxiv
Dual-Splitting Conformal Prediction for Multi-Step Time Series Forecasting
- Time series forecasting is crucial for applications like resource scheduling and risk management, where multi-step predictions provide a comprehensive view of future trends.
- The proposed Dual-Splitting Conformal Prediction (DSCP) method is a novel approach designed to capture inherent dependencies within time-series data for multi-step forecasting.
- Experimental results on real-world datasets demonstrate that DSCP outperforms existing Conformal Prediction methods, achieving a performance improvement of up to 23.59% compared to state-of-the-art techniques.
- DSCP is deployed in a real-world trajectory-based application for renewable energy generation and IT load forecasting, resulting in an 11.25% reduction in carbon emissions through predictive optimization of data center operations and controls.
Read Full Article
5 Likes
Arxiv
4d
223
Image Credit: Arxiv
HOT: Hadamard-based Optimized Training
- Researchers introduce a novel method called Hadamard-based Optimized Training (HOT) to optimize backpropagation in deep learning.
- HOT focuses on matrix multiplication, the most computationally expensive part of training, and applies Hadamard-based optimizations selectively.
- The method achieves up to 75% memory savings and a 2.6 times acceleration on GPUs, with minimal loss in accuracy compared to FP32 precision.
- HOT includes techniques such as Hadamard quantization, Hadamard low-rank approximation, activation buffer compression, and layer-wise quantizer selection.
Read Full Article
13 Likes
Arxiv
4d
264
Image Credit: Arxiv
Tricking Retrievers with Influential Tokens: An Efficient Black-Box Corpus Poisoning Attack
- Retrieval-augmented generation (RAG) systems, which incorporate external knowledge, are vulnerable to corpus poisoning attacks.
- Existing methods for crafting adversarial passages are slow and computationally expensive.
- A new method called Dynamic Importance-Guided Genetic Algorithm (DIGA) is proposed to efficiently generate adversarial passages.
- DIGA achieves superior efficiency and scalability, with comparable or better attack success rates.
Read Full Article
15 Likes
Arxiv
4d
366
Image Credit: Arxiv
Investigating the Duality of Interpretability and Explainability in Machine Learning
- The rapid evolution of machine learning has led to the widespread adoption of complex "black box" models.
- Efforts are focused on explaining these models instead of developing ones that are inherently interpretable.
- In this position paper, the imperative need for model interpretability is emphasized.
- An experimental evaluation of hybrid learning methods that integrate symbolic knowledge into neural network predictors is provided.
Read Full Article
22 Likes
Arxiv
4d
61
Image Credit: Arxiv
ProHOC: Probabilistic Hierarchical Out-of-Distribution Classification via Multi-Depth Networks
- ProHOC is a framework for detecting and classifying out-of-distribution (OOD) samples in a class hierarchy.
- The approach leverages a probabilistic model that uses networks trained for in-distribution (ID) classification at multiple hierarchy depths.
- Experiments conducted on three datasets with predefined class hierarchies demonstrate the effectiveness of the method.
- The code for ProHOC is available at https://github.com/walline/prohoc.
Read Full Article
3 Likes
Arxiv
4d
236
Image Credit: Arxiv
AdvSGM: Differentially Private Graph Learning via Adversarial Skip-gram Model
- The skip-gram model (SGM) is a popular graph embedding technique.
- However, the parameters of a released SGM may encode private information and pose privacy risks.
- AdvSGM is a differentially private skip-gram for graphs via adversarial training.
- Extensive experimental results on real-world graph datasets show that AdvSGM preserves high data utility.
Read Full Article
14 Likes
Arxiv
4d
362
Image Credit: Arxiv
Stochastic Engrams for Efficient Continual Learning with Binarized Neural Networks
- Researchers have proposed a novel approach for efficient continual learning in binarized neural networks.
- The approach integrates stochastically-activated engrams as a gating mechanism for metaplastic binarized neural networks (mBNNs).
- This method leverages the computational efficiency of mBNNs combined with the robustness of probabilistic memory traces to mitigate forgetting and maintain model reliability.
- The approach achieves high accuracies in class-incremental scenarios, comparable to state-of-the-art methods, while significantly reducing GPU and RAM usage.
Read Full Article
21 Likes
Arxiv
4d
215
Image Credit: Arxiv
DATA-WA: Demand-based Adaptive Task Assignment with Dynamic Worker Availability Windows
- Spatial crowdsourcing involves assigning location-based tasks to mobile workers.
- A new framework, Demand-based Adaptive Task Assignment with Dynamic Worker Availability Windows, is introduced to address the challenge of fluctuating demand and supply between tasks and workers over time.
- The framework includes task demand prediction and task assignment components.
- Experiments on real data show the effectiveness and efficiency of the proposed approach.
Read Full Article
12 Likes
Arxiv
4d
301
Image Credit: Arxiv
Adaptive Resampling with Bootstrap for Noisy Multi-Objective Optimization Problems
- The challenge of noisy multi-objective optimization lies in the constant trade-off between exploring new decision points and improving the precision of known points through resampling.
- This paper proposes a resampling decision function that incorporates the stochastic nature of the optimization problem by using bootstrapping and the probability of dominance.
- The approach utilizes bootstrap estimates of the means to achieve distribution-free estimation of the probability of dominance.
- The resampling approach is demonstrated to be efficient by applying it in the NSGA-II algorithm with a sequential resampling procedure under multiple noise variations.
Read Full Article
18 Likes
Arxiv
4d
236
Image Credit: Arxiv
F-INR: Functional Tensor Decomposition for Implicit Neural Representations
- F-INR is a framework that reformulates Implicit Neural Representation (INR) learning through functional tensor decomposition.
- It breaks down high-dimensional tasks into lightweight, axis-specific sub-networks, reducing computational costs.
- F-INR is modular, compatible with various INR architectures, and supports different decomposition modes.
- In experiments, F-INR trains 100 times faster than existing approaches while achieving higher fidelity in various tasks.
Read Full Article
14 Likes
Arxiv
4d
333
Image Credit: Arxiv
Uncertainty-aware Bayesian machine learning modelling of land cover classification
- Land cover classification involves the production of land cover maps using remote sensing imagery.
- A new Bayesian classification framework is proposed to incorporate input measurement uncertainty in land cover classification.
- The framework applies Bayesian quadratic discriminant analysis to land cover datasets from Copernicus Sentinel-2.
- The Bayesian models provide better interpretability, explicitly model input measurement uncertainty, and maintain predictive performance.
Read Full Article
20 Likes
For uninterrupted reading, download the app