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

A naukri.com initiative

New

Home

ML News

Arxiv

167

Image Credit: Arxiv

Contextualize-then-Aggregate: Circuits for In-Context Learning in Gemma-2 2B

In-Context Learning (ICL) is an intriguing ability of large language models (LLMs).
Research finds that Gemma-2 2B uses a two-step strategy, contextualize-then-aggregate, for task information assembly.
In the lower layers, the model builds up representations of individual fewshot examples, contextualized by preceding examples.
In the higher layers, these representations are aggregated to identify the task and prepare predictions.

Read Full Article

10 Likes

Arxiv

170

Image Credit: Arxiv

Data-driven Power Loss Identification through Physics-Based Thermal Model Backpropagation

A novel hybrid framework combining physics-based thermal modeling with data-driven techniques has been developed for accurate power loss identification in power electronics.
The framework leverages a cascaded architecture with a neural network that corrects the outputs of a nominal power loss model using temperature measurements.
Two neural architectures, a bootstrapped feedforward network and a recurrent neural network, were explored, with the feedforward approach achieving superior performance and computational efficiency.
Experimental results demonstrate that the hybrid model reduces temperature estimation errors and power loss prediction errors compared to traditional physics-based approaches, even in the presence of uncertainties.

Read Full Article

10 Likes

Arxiv

294

Image Credit: Arxiv

Towards Precise Action Spotting: Addressing Temporal Misalignment in Labels with Dynamic Label Assignment

Precise action spotting has attracted attention due to its applications.
Existing methods overlook a challenge of temporal misalignment in ground-truth labels.
A novel dynamic label assignment strategy is proposed to tackle this issue.
The method achieves state-of-the-art performance in conditions with temporal misalignment in labels.

Read Full Article

17 Likes

Arxiv

Image Credit: Arxiv

Nuclear Microreactor Control with Deep Reinforcement Learning

This study explores the application of deep reinforcement learning for real-time drum control in nuclear microreactors.
Deep reinforcement learning controllers demonstrate similar or better load-following performance compared to traditional PID control.
RL agents can reduce tracking error rate in short transients and maintain accuracy in longer, more complex load-following scenarios.
Multi-agent RL enables independent drum control and maintains reactor symmetry constraints without sacrificing performance.

Read Full Article

3 Likes

Discover more

Arxiv

241

Image Credit: Arxiv

Backdoor Detection through Replicated Execution of Outsourced Training

Outsourcing machine learning model training to cloud providers is common practice.
Detecting backdoored models without prior knowledge is challenging.
A client with access to multiple cloud providers can detect deviation by replicating training steps.
The approach is robust and suitable for clients with limited local compute capability.

Read Full Article

14 Likes

Arxiv

Image Credit: Arxiv

Self-Evolving Visual Concept Library using Vision-Language Critics

Researchers have introduced ESCHER, a visual concept library that aims to improve visual recognition.
ESCHER utilizes a vision-language model as a critic to iteratively refine the concept library.
The approach considers interactions between concepts and their impact on downstream classifiers.
ESCHER does not require human annotations and demonstrates effectiveness in various visual classification tasks.

Read Full Article

1 Like

Arxiv

Image Credit: Arxiv

Insight-RAG: Enhancing LLMs with Insight-Driven Augmentation

Retrieval Augmented Generation (RAG) frameworks enhance large language models (LLMs).
Insight-RAG is a framework designed to address limitations of conventional RAG methods.
Insight-RAG employs an LLM to analyze the query and extract informational requirements.
Integrating insight-driven retrieval in RAG enhances performance and expands applicability.

Read Full Article

2 Likes

Arxiv

Image Credit: Arxiv

Detecting Glioma, Meningioma, and Pituitary Tumors, and Normal Brain Tissues based on Yolov11 and Yolov8 Deep Learning Models

Accurate diagnosis of brain tumors is crucial for treatment planning and medical outcomes.
Manual interpretation of MRI scans is time-consuming and prone to errors.
Researchers propose using YoloV11 and YoloV8 deep learning models to detect glioma, meningioma, and pituitary brain tumors.
By fine-tuning the models, they achieve high accuracies and demonstrate the potential of CNNs in brain tumor detection.

Read Full Article

2 Likes

Arxiv

184

Image Credit: Arxiv

$\textit{Agents Under Siege}$: Breaking Pragmatic Multi-Agent LLM Systems with Optimized Prompt Attacks

Researchers have developed an adversarial attack that can bypass safety mechanisms in multi-agent Large Language Model (LLM) systems.
The attack optimizes prompt distribution across latency and bandwidth-constrained network topologies to maximize attack success rate while minimizing detection risk.
The method outperforms conventional attacks, exposing critical vulnerabilities in multi-agent systems.
Existing defenses, including variants of Llama-Guard and PromptGuard, fail to prohibit the attack.

Read Full Article

11 Likes

Arxiv

134

Image Credit: Arxiv

Rack Position Optimization in Large-Scale Heterogeneous Data Centers

This work presents a two-tier optimization framework for data center resource management in large-scale heterogeneous environments.
The framework combines deep reinforcement learning (DRL) with a gradient-based heuristic for optimal rack positioning.
The high-level DRL agent determines optimal rack type ordering, while the low-level heuristic minimizes movement counts and ensures fault-tolerant resource distribution.
The proposed approach outperformed the gradient-based heuristic and mixed-integer programming (MIP) solver in terms of objective value and computational efficiency.

Read Full Article

8 Likes

Arxiv

Image Credit: Arxiv

Aligning Diffusion Model with Problem Constraints for Trajectory Optimization

Diffusion models are effective for trajectory optimization but may violate critical constraints without explicit incorporation of constraint information.
A novel approach aligns diffusion models with problem-specific constraints using a hybrid loss function that measures and penalizes constraint violations during training.
The re-weighting strategy aligns predicted constraint violations to ground truth statistics, resulting in reduced violations compared to traditional diffusion models.
This approach can be integrated into the Dynamic Data-driven Application Systems (DDDAS) framework for efficient online trajectory adaptation.

Read Full Article

Arxiv

231

Image Credit: Arxiv

CopyQNN: Quantum Neural Network Extraction Attack under Varying Quantum Noise

Quantum Neural Networks (QNNs) have shown significant value across domains, with well-trained QNNs representing critical intellectual property often deployed via cloud-based QNN-as-a-Service (QNNaaS) platforms.
Existing approaches for QNN model extraction attacks have largely overlooked the impact of varying quantum noise inherent in noisy intermediate-scale quantum (NISQ) computers, limiting their effectiveness in real-world settings.
The CopyQNN framework proposes a three-step data cleaning method to eliminate noisy data based on its noise sensitivity, followed by the integration of contrastive and transfer learning within the quantum domain.
Experimental results on NISQ computers demonstrate that the practical implementation of CopyQNN outperforms state-of-the-art QNN extraction attacks, achieving an average performance improvement of 8.73% while reducing the number of required queries by 90x.

Read Full Article

13 Likes

Arxiv

315

Image Credit: Arxiv

Spatiotemporal Attention Learning Framework for Event-Driven Object Recognition

A novel spatiotemporal learning framework for event-based object recognition is presented.
The framework utilizes a VGG network enhanced with Convolutional Block Attention Module (CBAM).
The approach achieves comparable performance to state-of-the-art ResNet-based methods while reducing parameter count.
Experimental results highlight the efficiency and effectiveness of the framework for real-world applications.

Read Full Article

18 Likes

Arxiv

110

Image Credit: Arxiv

Using complex prompts to identify fine-grained biases in image generation through ChatGPT-4o

Large AI models can reveal both bias in training data or AI output, and bias in society, such as social disparities.
Social biases often appear in image generation AI outputs as 'marked' features.
Generative AI is sensitive to marked features, often exacerbating social biases.
Complex prompts can be used to investigate bias in image generation AI models.

Read Full Article

6 Likes

Arxiv

Image Credit: Arxiv

Flow Matching on Lie Groups

Flow Matching (FM) is a generative modelling technique that aims to learn how to sample from a target distribution by flowing samples from an easy-to-sample distribution.
Existing FM methods use straight line segments or geodesics for flow matching, but they are only well-defined on Euclidean space or Riemannian manifolds.
The alternative approach in this work is to apply FM on Lie groups by using exponential curves instead of straight line segments, leading to a simple and fast implementation using matrix operations.
FM on Lie groups can be used for generative modelling with data consisting of sets of features and poses.

Read Full Article

4 Likes

For uninterrupted reading, download the app