A naukri.com initiative
ML News
Arxiv
2d
206
Image Credit: Arxiv
Exploring Data Scaling Trends and Effects in Reinforcement Learning from Human Feedback
- Reinforcement Learning from Human Feedback (RLHF) is crucial for aligning large language models with human preferences.
- This paper explores data-driven bottlenecks in RLHF performance scaling, focusing on reward hacking and decreasing response diversity.
- The hybrid reward system, combining reasoning task verifiers (RTV) and a generative reward model (GenRM), is introduced to mitigate reward hacking.
- The novel prompt-selection method, Pre-PPO, is proposed to maintain response diversity and enhance learning effectiveness.
Read Full Article
12 Likes
Arxiv
2d
226
Image Credit: Arxiv
WeatherMesh-3: Fast and accurate operational global weather forecasting
- WeatherMesh-3 (WM-3) is a transformer-based global weather forecasting system that improves accuracy and computational efficiency.
- It introduces a latent rollout for arbitrary-length predictions in latent space and a modular architecture for blended initial conditions.
- WM-3 generates 14-day global forecasts at 0.25-degree resolution in 12 seconds on a single RTX 4090, achieving superior accuracy and significant speedup.
- This model aims to democratize weather forecasting and push the performance boundaries of machine learning-based prediction.
Read Full Article
13 Likes
Arxiv
2d
83
Image Credit: Arxiv
DynaGraph: Interpretable Multi-Label Prediction from EHRs via Dynamic Graph Learning and Contrastive Augmentation
- DynaGraph is an interpretable contrastive graph model that learns the dynamics of multivariate time-series electronic health records (EHRs) as part of optimization.
- DynaGraph captures the hidden dependencies of the multivariate time-series using dynamic graph construction.
- Compared to state-of-the-art models, DynaGraph achieves significant improvements in accuracy and sensitivity in various clinical datasets.
- DynaGraph provides clinical validation by indicating the importance of clinical covariates over time.
Read Full Article
5 Likes
Arxiv
2d
190
Image Credit: Arxiv
FLIP: Towards Comprehensive and Reliable Evaluation of Federated Prompt Learning
- The paper introduces a framework called FLIP to evaluate federated prompt learning algorithms.
- FLIP assesses the performance of 8 state-of-the-art federated prompt learning methods across various scenarios.
- The findings demonstrate that prompt learning maintains strong generalization performance with minimal resource consumption.
- This work emphasizes the effectiveness of federated prompt learning in data scarcity and cross-domain distributional shift scenarios.
Read Full Article
11 Likes
Arxiv
2d
325
Image Credit: Arxiv
Machine Learning Models for Soil Parameter Prediction Based on Satellite, Weather, Clay and Yield Data
- The AgroLens project aims to develop Machine Learning-based methodologies for predicting soil nutrient levels without relying on laboratory tests.
- The project utilizes the LUCAS Soil dataset and Sentinel-2 satellite imagery to estimate key soil properties including phosphorus, potassium, nitrogen, and pH levels.
- Supplementary features like weather data, harvest rates, and Clay AI-generated embeddings are integrated to enhance the soil nutrient prediction model.
- Implementation of advanced algorithms such as Random Forests, Extreme Gradient Boosting (XGBoost), and Fully Connected Neural Networks (FCNN) results in robust model performance and accurate nutrient prediction.
Read Full Article
19 Likes
Arxiv
2d
214
Image Credit: Arxiv
Hybrid Time-Domain Behavior Model Based on Neural Differential Equations and RNNs
- Researchers propose a novel continuous-time domain hybrid modeling paradigm for nonlinear dynamics system identification.
- The hybrid model integrates neural network differential models with recurrent neural networks (RNNs).
- Theoretical analysis demonstrates advantages in event-driven dynamic mutation response and gradient propagation stability.
- Validation using real data shows improved fitting accuracy and effectiveness in capturing nonlinear memory effects.
Read Full Article
12 Likes
Arxiv
2d
305
Image Credit: Arxiv
MASCOTS: Model-Agnostic Symbolic COunterfactual explanations for Time Series
- Counterfactual explanations provide an intuitive way to understand model decisions by identifying minimal changes required to alter an outcome.
- MASCOTS is a method that leverages the Bag-of-Receptive-Fields representation alongside symbolic transformations to enhance interpretability while preserving fidelity to the original data and model.
- MASCOTS directly generates meaningful and diverse counterfactual observations in a model-agnostic manner, operating on both univariate and multivariate data.
- MASCOTS improves interpretability and sparsity, allowing for explanations in visual, natural language, or semantic representations.
Read Full Article
18 Likes
Arxiv
2d
19
Image Credit: Arxiv
On-site estimation of battery electrochemical parameters via transfer learning based physics-informed neural network approach
- This paper presents a novel physical parameter estimation framework for on-site model characterization, using a two-phase modelling strategy with Physics-Informed Neural Networks (PINNs) and transfer learning (TL).
- In the first phase, a PINN is trained using only the physical principles of the single particle model (SPM) equations.
- In the second phase, the majority of the PINN parameters are frozen, while critical electrochemical parameters are set as trainable and adjusted using real-world voltage profile data.
- The proposed approach significantly reduces computational costs, making it suitable for real-time implementation on Battery Management Systems (BMS).
Read Full Article
1 Like
Arxiv
2d
325
Image Credit: Arxiv
Generative Reliability-Based Design Optimization Using In-Context Learning Capabilities of Large Language Models
- Large Language Models (LLMs) have in-context learning capabilities for reasoning, problem-solving, and pattern recognition.
- A generative design method is proposed that combines LLMs with metaheuristic algorithms for reliability-based design optimization.
- Reliability analysis is performed using LLMs and Kriging surrogate modeling to reduce computational burden.
- Experimental results demonstrate the effectiveness of the proposed approach in identifying feasible solutions.
Read Full Article
19 Likes
Arxiv
2d
146
Image Credit: Arxiv
STADE: Standard Deviation as a Pruning Metric
- Large Language Models (LLMs) are widely used for various tasks, but they require long training times and large model sizes.
- Pruning methods like Wanda can reduce computational demands without retraining and are effective in maintaining performance.
- This study provides a theoretical explanation of the effectiveness of Wanda and introduces a new pruning method called STADE based on the standard deviation of the input.
- Experiments on Llama and Open Pre-trained Transformers (OPT) models validate the theoretical findings, demonstrating the variability of Wanda's optimal performance depending on training conditions.
Read Full Article
8 Likes
Arxiv
2d
158
Image Credit: Arxiv
A Causal Framework to Measure and Mitigate Non-binary Treatment Discrimination
- Fairness analyses of algorithmic decision-making should include non-binary treatment decisions.
- A causal framework is proposed to measure treatment disparity and its impact on outcomes.
- Loan approval datasets are analyzed to reveal potential discrimination in non-binary treatment decisions.
- The framework can mitigate treatment discrimination and ensure fair decision-making processes.
Read Full Article
9 Likes
Arxiv
2d
11
Image Credit: Arxiv
DeepOFormer: Deep Operator Learning with Domain-informed Features for Fatigue Life Prediction
- Researchers propose DeepOFormer, a deep operator learning framework for fatigue life prediction.
- It addresses the challenge of overfitting using a transformer-based encoder and a mean L2 relative error loss function.
- Domain-informed features, such as Stussi, Weibull, and Pascual and Meeker (PM), are considered to improve prediction accuracy.
- DeepOFormer achieves superior performance compared to state-of-the-art deep/machine learning methods in predicting fatigue life in aluminum alloys.
Read Full Article
Like
Arxiv
2d
115
Image Credit: Arxiv
Probabilistic Uncertain Reward Model: A Natural Generalization of Bradley-Terry Reward Model
- Reinforcement Learning from Human Feedback (RLHF) has emerged as a critical technique for training large language models.
- The Probabilistic Uncertain Reward Model (PURM) is proposed as a natural generalization of the classical Bradley-Terry reward model.
- PURM learns reward distributions directly from preference data and quantifies per-sample uncertainty.
- Experiments demonstrate that PURM significantly delays reward hacking and improves final reward performance.
Read Full Article
6 Likes
Arxiv
2d
134
Image Credit: Arxiv
SPDNet: Seasonal-Periodic Decomposition Network for Advanced Residential Demand Forecasting
- Residential electricity demand forecasting is important for efficient energy management and grid stability.
- A novel deep learning framework called SPDNet is proposed to tackle the challenges of capturing intricate temporal dynamics in electricity demand data.
- SPDNet consists of two main modules: Seasonal-Trend Decomposition Module (STDM) and Periodical Decomposition Module (PDM).
- Extensive experiments show that SPDNet outperforms traditional and advanced models in both accuracy and efficiency.
Read Full Article
8 Likes
Arxiv
2d
269
Image Credit: Arxiv
Learnable cut flow
- Neural networks have become popular in high energy physics but their training process is opaque.
- Learnable Cut Flow (LCF) is a neural network that combines the simplicity of the traditional cut flow method with the power of data-driven training.
- LCF implements two cut strategies, parallel and sequential, to determine optimal boundaries for cut selection.
- LCF offers insights into feature importance and performs effectively in real-world scenarios.
Read Full Article
16 Likes
For uninterrupted reading, download the app