menu
techminis

A naukri.com initiative

google-web-stories
Home

>

ML News

ML News

source image

Marktechpost

1d

read

159

img
dot

Sequential-NIAH: A Benchmark for Evaluating LLMs in Extracting Sequential Information from Long Texts

  • Evaluating how well LLMs handle long contexts is essential, especially for retrieving specific, relevant information embedded in lengthy inputs.
  • Needle-in-a-Haystack (NIAH) task challenges models to retrieve critical information from predominantly irrelevant content and lacks tasks involving retrieval and correct ordering of sequential information.
  • Sequential-NIAH benchmark designed to assess how well LLMs retrieve sequential information, referred to as a needle, from long texts.
  • Tests on popular LLMs showed highest performance at just 63.15%, highlighting the difficulty of the task and need for further advancement.

Read Full Article

like

9 Likes

source image

Medium

1d

read

572

img
dot

Image Credit: Medium

Why AI Still Can’t Get Sarcasm ?

  • AI struggles to comprehend sarcasm as it cannot understand the subtextual cues and contextual shifts.
  • Sarcasm requires understanding the intention behind the words, which is challenging for AI.
  • AI language models may improve in detecting tone and even being sarcastic, but they will never experience true empathy.
  • Sarcasm remains a human tool for social commentary and a defense mechanism that machines cannot fully replicate.

Read Full Article

like

10 Likes

source image

Mit

2d

read

237

img
dot

Image Credit: Mit

Robotic system zeroes in on objects most relevant for helping humans

  • MIT roboticists have developed a system, called "Relevance," to help robots focus on relevant features in a scene for assisting humans.
  • The Relevance approach enables robots to determine a human's objective using cues like audio and visual information.
  • A robot can then identify objects most likely to be relevant in fulfilling the human's objective and act accordingly.
  • In an experiment simulating a conference breakfast buffet, the robot successfully assisted humans in various scenarios with high accuracy.
  • The robot predicted a human's objective with 90% accuracy and identified relevant objects with 96% accuracy.
  • This method not only improves a robot's efficiency but also enhances safety by reducing collisions by over 60%.
  • The system mimics the human brain's Reticular Activating System to selectively process and filter information.
  • It consists of phases like perception, trigger check, relevance determination, and object offering based on relevance.
  • The researchers aim to apply this system in smart manufacturing, warehouse environments, and household tasks for more natural human-robot interactions.
  • The team's goal is to enable robots to offer seamless, intelligent, safe, and efficient assistance in dynamic environments.

Read Full Article

like

14 Likes

source image

Arxiv

2d

read

120

img
dot

Image Credit: Arxiv

Representation Learning for Tabular Data: A Comprehensive Survey

  • Tabular data is prevalent in machine learning classification and regression applications.
  • Deep Neural Networks (DNNs) are promising for representation learning in tabular data.
  • Existing models for tabular representation learning are categorized into specialized, transferable, and general models.
  • The survey also covers ensemble methods and extensions of tabular learning.

Read Full Article

like

7 Likes

source image

Arxiv

2d

read

82

img
dot

Image Credit: Arxiv

Active Learning Methods for Efficient Data Utilization and Model Performance Enhancement

  • This paper gives an overview of Active Learning (AL) in machine learning, which helps models achieve better performance using fewer labeled examples.
  • AL is used in various fields such as computer vision, natural language processing, transfer learning, and real-world applications.
  • The paper focuses on topics like uncertainty estimation, handling class imbalance, fairness, and creation of strong evaluation metrics.
  • AL often gives better results than passive learning, especially with good evaluation measures.

Read Full Article

like

4 Likes

source image

Arxiv

2d

read

179

img
dot

Image Credit: Arxiv

SparseJEPA: Sparse Representation Learning of Joint Embedding Predictive Architectures

  • SparseJEPA is an extension of Joint Embedding Predictive Architectures (JEPA) that integrates sparse representation learning to enhance the quality of learned representations.
  • SparseJEPA encourages shared latent space variables among data features with strong semantic relationships while maintaining predictive performance.
  • The architecture was tested on the CIFAR-100 dataset and a lightweight Vision Transformer for image classification and low-level tasks using transfer learning.
  • Incorporating sparsity improves the latent space, leading to more meaningful and interpretable representations.

Read Full Article

like

10 Likes

source image

Arxiv

2d

read

182

img
dot

Image Credit: Arxiv

Deep Learning Meets Process-Based Models: A Hybrid Approach to Agricultural Challenges

  • Process-based models (PBMs) and deep learning (DL) are two key approaches in agricultural modelling.
  • This study presents a review of PBMs, DL models, and hybrid PBM-DL frameworks in agricultural modelling.
  • Results demonstrate that hybrid models consistently outperform traditional PBMs and DL models.
  • The study contributes to the development of scalable, interpretable, and reproducible agricultural models for sustainable agriculture.

Read Full Article

like

10 Likes

source image

Arxiv

2d

read

70

img
dot

Image Credit: Arxiv

Hexcute: A Tile-based Programming Language with Automatic Layout and Task-Mapping Synthesis

  • Hexcute is a tile-based programming language designed for deep learning workloads.
  • It exposes shared memory and register abstractions to enable fine-grained optimization.
  • Hexcute automates layout and task mapping synthesis with a type-inference-based algorithm.
  • Evaluation shows that Hexcute achieves significant speedup over existing DL compilers.

Read Full Article

like

4 Likes

source image

Arxiv

2d

read

326

img
dot

Image Credit: Arxiv

General Post-Processing Framework for Fairness Adjustment of Machine Learning Models

  • A new framework for fairness adjustments in machine learning models has been introduced in a recent paper on arXiv.
  • The framework applies to various machine learning tasks, including regression and classification, and supports different fairness metrics.
  • Unlike traditional approaches, this method adapts in-processing techniques for post-processing, providing greater flexibility in model development.
  • The advantages of this framework include preserving model performance, eliminating the need for custom loss functions, accommodating black-box systems, and providing interpretable insights.

Read Full Article

like

19 Likes

source image

Arxiv

2d

read

167

img
dot

Image Credit: Arxiv

FairPlay: A Collaborative Approach to Mitigate Bias in Datasets for Improved AI Fairness

  • FairPlay is a web-based software application that enables multiple stakeholders to collaboratively debias datasets.
  • Users can negotiate and arrive at a mutually acceptable outcome without a universally agreed-upon theory of fairness.
  • Through user studies, FairPlay has demonstrated the ability to help users reach a consensus within about five rounds of gameplay.
  • The application has the potential to enhance fairness in AI systems by addressing the issue of bias in datasets.

Read Full Article

like

10 Likes

source image

Arxiv

2d

read

276

img
dot

Image Credit: Arxiv

Learning Energy-Based Generative Models via Potential Flow: A Variational Principle Approach to Probability Density Homotopy Matching

  • Energy-based models (EBMs) are a powerful class of probabilistic generative models.
  • Variational Potential Flow Bayes (VPFB) is a new energy-based generative framework that eliminates the need for implicit MCMC sampling.
  • VPFB learns an energy-parameterized potential flow by matching a flow-driven density homotopy to the data distribution.
  • Experimental results show that VPFB performs competitively with existing approaches in terms of sample quality and versatility.

Read Full Article

like

16 Likes

source image

Arxiv

2d

read

281

img
dot

Image Credit: Arxiv

Gradient-Optimized Fuzzy Classifier: A Benchmark Study Against State-of-the-Art Models

  • A benchmark study was conducted to compare the performance of a Gradient-Optimized Fuzzy Inference System (GF) classifier against various state-of-the-art machine learning models.
  • The evaluation was done on five datasets from the UCI Machine Learning Repository, each offering diverse input types, class distributions, and classification complexity.
  • Compared to traditional Fuzzy Inference Systems, the GF classifier, which utilizes gradient descent, demonstrated superior classification accuracy, high precision, and significantly lower training times.
  • The GF model exhibited consistency across folds and datasets, indicating its robustness in handling noisy data and varying feature sets.

Read Full Article

like

16 Likes

source image

Arxiv

2d

read

299

img
dot

Image Credit: Arxiv

Boosting Classifier Performance with Opposition-Based Data Transformation

  • A novel data transformation framework based on Opposition-Based Learning (OBL) is introduced to boost the performance of traditional classification algorithms.
  • OBL generates synthetic opposite samples to replace the training data and improve decision boundary formation.
  • Three OBL variants, Global OBL, Class-Wise OBL, and Localized Class-Wise OBL, are explored and integrated with popular classifiers.
  • Experiments show that OBL-enhanced classifiers consistently outperform standard counterparts in terms of accuracy and F1-score, particularly in complex or sparse learning environments.

Read Full Article

like

18 Likes

source image

Arxiv

2d

read

7

img
dot

Image Credit: Arxiv

Learning Explainable Dense Reward Shapes via Bayesian Optimization

  • Current reinforcement learning from human feedback (RLHF) pipelines for large language model (LLM) alignment typically assign scalar rewards to sequences, using the final token as a surrogate indicator for the quality of the entire sequence.
  • This work proposes a reward-shaping function that leverages explainability methods such as SHAP and LIME to estimate per-token rewards from the reward model.
  • The study utilizes a bilevel optimization framework that integrates Bayesian Optimization and policy training to handle noise from token reward estimates.
  • Experiments demonstrate that achieving a better balance of token-level reward attribution leads to improved performance and faster training of the optimal policy.

Read Full Article

like

Like

source image

Arxiv

2d

read

19

img
dot

Image Credit: Arxiv

Quantum Doubly Stochastic Transformers

  • Researchers have developed a hybrid classical-quantum doubly stochastic Transformer (QDSFormer) which incorporates a variational quantum circuit in place of the Softmax in the self-attention layer.
  • The QDSFormer yields more diverse doubly stochastic matrices (DSMs) that better preserve information compared to classical operators.
  • In multiple small-scale object recognition tasks, the QDSFormer outperforms a standard Vision Transformer and other doubly stochastic Transformers.
  • The QDSFormer shows improved training stability and lower performance variation, potentially mitigating the unstable training of Vision Transformers on small-scale data.

Read Full Article

like

1 Like

For uninterrupted reading, download the app