Machine Learning (ML) Latest News and Trending articles from all top sources only on Techminis

A naukri.com initiative

New

Home

ML News

Arxiv

Image Credit: Arxiv

Leveraging Generative Adversarial Networks for Addressing Data Imbalance in Financial Market Supervision

This study explores the application of generative adversarial networks in financial market supervision to address data imbalance and improve risk prediction accuracy.
Traditional models struggle to identify minority events in imbalanced financial market data.
The study proposes using GAN to generate synthetic data that resembles these minority events.
Experimental results indicate that GAN-generated data outperforms traditional oversampling and undersampling methods, showing potential for use in financial regulatory agencies.

Read Full Article

Arxiv

Image Credit: Arxiv

Multi-Branch Mutual-Distillation Transformer for EEG-Based Seizure Subtype Classification

Researchers propose Multi-Branch Mutual-Distillation (MBMD) Transformer for EEG-based seizure subtype classification.
MBMD Transformer is designed to effectively train from small labeled data.
It replaces even-numbered encoder blocks of the vanilla Vision Transformer with multi-branch encoder blocks.
Experiments show that MBMD Transformer outperforms traditional machine learning and deep learning approaches for EEG-based seizure subtype classification.

Read Full Article

Arxiv

266

Image Credit: Arxiv

Dipper: Diversity in Prompts for Producing Large Language Model Ensembles in Reasoning tasks

Large Language Models still encounter challenges in reasoning tasks, especially for smaller models.
Inference-time methods like prompting have shown effectiveness but rely on sequential queries.
The ensemble method, running multiple models in parallel, is a promising approach for better inference-time performance.
A novel training-free LLM ensemble framework is proposed for improved reasoning in math tasks.

Read Full Article

16 Likes

Arxiv

159

Image Credit: Arxiv

MPPO: Multi Pair-wise Preference Optimization for LLMs with Arbitrary Negative Samples

MPPO is a new algorithm for preference optimization in large language models (LLMs) with arbitrary negative samples.
Existing methods like DPO and KTO rely heavily on abundant preference data and require a reference model.
MPPO leverages the average likelihood of model responses to fit the reward function, maximizing the utilization of preference data.
Experimental results show that MPPO outperforms other methods like DPO and ORPO across various benchmarks.

Read Full Article

9 Likes

Discover more

Arxiv

139

Image Credit: Arxiv

LLMs for Literature Review: Are we there yet?

Literature reviews are time-intensive and challenging to write, but recent Large Language Models (LLMs) can assist in the process.
The task is decomposed into two components: retrieving related works based on an abstract and writing the literature review.
LLMs are effective in retrieving relevant papers using a two-step search strategy and a prompting-based re-ranking mechanism.
A two-step approach, involving planning and execution, improves the quality of the generated literature reviews.

Read Full Article

8 Likes

Arxiv

Image Credit: Arxiv

An Enhanced Text Compression Approach Using Transformer-based Language Models

Text compression is important for shrinking textual data and conserving storage, bandwidth, and computational efficiency.
A new method called RejuvenateForme using transformer-based approach is proposed for text decompression.
The method incorporates a meticulous pre-processing technique and a lossless compression algorithm.
RejuvenateForme achieves state-of-the-art compression ratios and shows comprehensive efficacy in terms of BLEU scores.

Read Full Article

3 Likes

Arxiv

Image Credit: Arxiv

Structured Extraction of Real World Medical Knowledge using LLMs for Summarization and Search

Creation and curation of knowledge graphs can accelerate disease discovery and analysis in real-world data.
Proposes creating patient knowledge graphs using large language model extraction techniques, allowing data extraction via natural language rather than rigid ontological hierarchies.
Demonstrates the method through patient search for Dravet syndrome using a large ambulatory care EHR database.
Applies the method to identify Beta-propeller protein-associated neurodegeneration (BPAN) patients, showing real-world discovery where no ground truth exists.

Read Full Article

1 Like

Arxiv

298

Image Credit: Arxiv

Toxicity Detection towards Adaptability to Changing Perturbations

Toxicity detection is crucial for maintaining the peace of the society.
Existing methods are vulnerable to evolving perturbation patterns created by malicious users.
A new dataset with 9 types of perturbation patterns has been constructed to validate the vulnerability of current methods.
A domain incremental learning paradigm and benchmark are proposed to ensure robustness to dynamically emerging types of perturbed toxic text.

Read Full Article

17 Likes

Arxiv

290

Image Credit: Arxiv

The Reliability Paradox: Exploring How Shortcut Learning Undermines Language Model Calibration

Recent studies have found pre-trained language models (PLMs) to suffer from miscalibration, indicating a lack of accuracy in confidence estimates.
Evaluation methods assuming lower calibration error estimates indicate more reliable predictions may be flawed.
Fine-tuned PLMs often resort to shortcuts, leading to overconfident predictions that lack generalizability.
Models with seemingly superior calibration actually have higher levels of non-generalizable decision rules.

Read Full Article

17 Likes

Arxiv

179

Image Credit: Arxiv

Exploring Query Efficient Data Generation towards Data-free Model Stealing in Hard Label Setting

Data-free model stealing involves replicating the functionality of a target model into a substitute model without accessing the target model's structure, parameters, or training data.
Existing methods within cooperative game frameworks often produce samples with high confidence for the prediction of the substitute model, which makes it difficult for the substitute model to replicate the behavior of the target model.
This paper presents a new data-free model stealing approach called Query Efficient Data Generation (QEDG).
The proposed method achieved better performance with fewer queries compared to the state-of-the-art methods on real MLaaS scenario and five datasets.

Read Full Article

10 Likes

Arxiv

167

Image Credit: Arxiv

Functional connectomes of neural networks

The study of the functional connectome, which maps the functional connections between different brain regions, has provided valuable insights through various advanced analysis techniques developed over the years.
Neural networks, inspired by the brain's architecture, have achieved notable success in diverse applications but are often noted for their lack of interpretability.
A novel approach is proposed in this paper to bridge neural networks and human brain functions by leveraging brain-inspired techniques.
The approach enhances the interpretability of neural networks, providing a deeper understanding of their underlying mechanisms.

Read Full Article

10 Likes

Arxiv

Image Credit: Arxiv

Channel Merging: Preserving Specialization for Merged Experts

The practice of utilizing task-specific fine-tuning has been implemented to improve the performance of large language models (LLM) in subsequent tasks.
Model merging strategies have emerged to reduce the memory footprint during inference by merging all LLMs into one model.
To mitigate parameter conflicts and improve storage efficiency, a novel strategy called Channel Merging is introduced.
Channel Merging clusters and merges channel parameters based on their similarity, reducing conflicts while preserving specialized knowledge.

Read Full Article

4 Likes

Arxiv

Image Credit: Arxiv

Maximize Your Data's Potential: Enhancing LLM Accuracy with Two-Phase Pretraining

Pretraining large language models effectively requires strategic data selection, blending and ordering.
A two-phase pretraining approach outperforms random data ordering and natural distribution of tokens.
The two-phase approach improves average accuracies by 3.4% and 17%.
Guidance is provided on crafting optimal data blends based on data source quality and the number of epochs.

Read Full Article

3 Likes

Arxiv

Image Credit: Arxiv

Inference-Aware Fine-Tuning for Best-of-N Sampling in Large Language Models

Researchers propose a novel inference-aware fine-tuning paradigm for large language models (LLMs).
The paradigm focuses on optimizing the performance of the inference-time strategy.
Imitation learning and reinforcement learning methods are devised to tackle the non-differentiable argmax operator within the Best-of-N (BoN) inference strategy.
Experiments show improved performance and inference-time compute using BoN-aware fine-tuning.

Read Full Article

3 Likes

Arxiv

346

Image Credit: Arxiv

Deep reinforcement learning with time-scale invariant memory

Researchers integrate a computational neuroscience model of scale invariant memory into deep reinforcement learning (RL) agents.
Agents built with scale invariant memory can learn robustly across a wide range of temporal scales, unlike agents built with commonly used recurrent memory architectures such as LSTM.
This integration of computational principles from neuroscience and cognitive science enhances adaptability to complex temporal dynamics in deep neural networks.
The result mirrors some of the core properties of human learning.

Read Full Article

20 Likes

For uninterrupted reading, download the app