A naukri.com initiative
ML News
Arxiv
1w
42
Image Credit: Arxiv
Controller Distillation Reduces Fragile Brain-Body Co-Adaptation and Enables Migrations in MAP-Elites
- Brain-body co-optimization suffers from fragile co-adaptation, hindering transferability.
- Previous approach using MAP-Elites still suffers from disrupted co-adaptation.
- Utilizing 'Pollination' technique helps reduce fragile co-adaptation and promotes migrations in MAP-Elites.
- Pollination increases success of body mutations and improves quality-diversity metrics.
Read Full Article
2 Likes
Arxiv
1w
186
Image Credit: Arxiv
Lugha-Llama: Adapting Large Language Models for African Languages
- Large language models (LLMs) struggle to recognize low-resource languages, including African languages.
- Combining curated data from African languages with high-quality English educational texts significantly improves the model's performance on these languages.
- On the IrokoBench dataset, the models consistently achieve the best performance compared to other baselines, particularly on knowledge-intensive multiple-choice questions (AfriMMLU).
- The models outperform the base model by over 10% on the cross-lingual question answering benchmark AfriQA.
Read Full Article
11 Likes
Arxiv
1w
156
Image Credit: Arxiv
Diffusion Factor Models: Generating High-Dimensional Returns with Factor Structure
- Researchers propose a diffusion factor model that integrates latent factor structure into generative diffusion processes for financial scenario simulation.
- By exploiting the low-dimensional factor structure inherent in asset returns, the model decomposes the score function using time-varying orthogonal projections.
- The proposed model provides nonasymptotic error bounds for both score estimation and generated distribution, surpassing the dimension-dependent limits in classical nonparametric statistics literature.
- Numerical studies confirm superior performance in latent subspace recovery and empirical analysis demonstrates the economic significance of the framework.
Read Full Article
9 Likes
Arxiv
1w
410
Image Credit: Arxiv
CAFE-AD: Cross-Scenario Adaptive Feature Enhancement for Trajectory Planning in Autonomous Driving
- Imitation learning based planning tasks on the nuPlan dataset have gained interest in generating human-like driving behaviors.
- To tackle challenges in closed-loop testing and long-tail distribution of scenarios, CAFE-AD method is introduced for trajectory planning in autonomous driving.
- CAFE-AD includes an adaptive feature pruning module to capture relevant information and reduce noisy interference.
- The cross-scenario feature interpolation module enhances scenario information and alleviates over-fitting in dominant scenarios.
Read Full Article
24 Likes
Arxiv
1w
63
Image Credit: Arxiv
InteractRank: Personalized Web-Scale Search Pre-Ranking with Cross Interaction Features
- Modern search systems use a multi-stage architecture to deliver personalized results efficiently.
- The pre-ranking stage, vital for scoring and filtering items, often lacks in capturing complex interactions.
- InteractRank is a novel two tower pre-ranking model with robust cross interaction features used at Pinterest.
- In real-world experiments, InteractRank improves online engagement metric by 6.5% over a BM25 baseline and 3.7% over a vanilla two tower baseline.
Read Full Article
3 Likes
Arxiv
1w
317
Image Credit: Arxiv
FACT: Multinomial Misalignment Classification for Point Cloud Registration
- The paper introduces FACT, a method for predicting alignment quality of registered lidar point cloud pairs.
- FACT extracts local features from a registered pair and processes them with a point transformer-based network to predict a misalignment class.
- FACT outperforms both direct regression and prior binary classification by introducing a custom regression-by-classification loss function.
- The method successfully classifies point-cloud pairs registered with ICP and GeoTransformer, and can assist in correcting misaligned point-cloud maps.
Read Full Article
19 Likes
Arxiv
1w
405
Image Credit: Arxiv
SEE: Continual Fine-tuning with Sequential Ensemble of Experts
- Continual fine-tuning of large language models (LLMs) suffers from catastrophic forgetting.
- The Sequential Ensemble of Experts (SEE) framework is introduced to address the challenges of continual fine-tuning.
- SEE allows each expert to independently decide whether a query should be handled, removing the need for an additional router.
- Experiments reveal that SEE outperforms prior approaches in continual fine-tuning and demonstrates remarkable generalization ability.
Read Full Article
24 Likes
Arxiv
1w
126
Image Credit: Arxiv
Learning global control of underactuated systems with Model-Based Reinforcement Learning
- This paper presents a solution for the AI Olympics competition held at ICRA 2025.
- The solution utilizes the MC-PILCO algorithm, known for its data efficiency in low-dimensional robotic tasks.
- MC-PILCO optimizes a system dynamics model using interaction data for policy refinement through simulation.
- The algorithm has previously won the first two editions of the competition, demonstrating its effectiveness in both simulated and real-world environments.
Read Full Article
7 Likes
Arxiv
1w
372
Image Credit: Arxiv
Hybrid machine learning models based on physical patterns to accelerate CFD simulations: a short guide on autoregressive models
- Researchers have developed hybrid machine learning models for fluid flow simulations.
- The models integrate High-Order Singular Value Decomposition (HOSVD) with Long Short-Term Memory (LSTM) architectures.
- HOSVD improves dimensionality reduction by preserving multidimensional structures.
- The models demonstrate improved accuracy and reliability in predicting complex fluid dynamics.
Read Full Article
22 Likes
Arxiv
1w
202
Image Credit: Arxiv
Domain-Specific Pruning of Large Mixture-of-Experts Models with Few-shot Demonstrations
- Mixture-of-Experts models achieve performance and inference efficiency by activating only a subset of experts.
- Large-scale Mixture-of-Experts models face the limitation of storing all experts which leads to significant memory overhead.
- A pruning framework called EASY-EP is proposed, which utilizes domain-specific demonstrations to identify and retain the most relevant experts.
- EASY-EP can achieve comparable performance and higher throughput while reducing memory usage by half in the DeepSeek-R1 model.
Read Full Article
12 Likes
Arxiv
1w
253
Image Credit: Arxiv
Hybrid CNN with Chebyshev Polynomial Expansion for Medical Image Analysis
- Automated detection of pulmonary nodules in CT scans is challenging due to variability in nodule characteristics.
- Traditional CNNs have limitations in capturing fine-grained variations in medical images.
- A new hybrid approach, Chebyshev-CNN, integrates Chebyshev polynomial expansions into CNN layers to improve representation of anatomical structures.
- The Chebyshev-CNN model achieves superior performance in classifying pulmonary nodules and shows potential for broader applications in clinical decision support systems.
Read Full Article
15 Likes
Arxiv
1w
283
Image Credit: Arxiv
Deep Neural Koopman Operator-based Economic Model Predictive Control of Shipboard Carbon Capture System
- Shipboard carbon capture is a promising solution to reduce carbon emissions in international shipping.
- A data-driven dynamic modeling and economic predictive control approach is proposed within the Koopman framework for shipboard post-combustion carbon capture plants.
- A deep neural Koopman operator modeling approach is used to establish a time-varying model predicting economic operational cost and system outputs based on accessible state measurements.
- The proposed method improves economic operational performance, carbon capture rate, and ensures safe operation by satisfying hard constraints on system outputs.
Read Full Article
17 Likes
Arxiv
1w
123
Image Credit: Arxiv
ZIP: An Efficient Zeroth-order Prompt Tuning for Black-box Vision-Language Models
- Recent studies have introduced various approaches for prompt-tuning black-box vision-language models, referred to as black-box prompt-tuning (BBPT).
- To address the issue of excessive queries in prompt-tuning, a new approach called Zeroth-order Intrinsic-dimensional Prompt-tuning (ZIP) is proposed.
- ZIP reduces problem dimensionality and variance of zeroth-order gradient estimates for efficient and robust prompt optimization.
- ZIP achieves state-of-the-art performance on multiple vision-language tasks, with improved few-shot accuracy and query efficiency.
Read Full Article
7 Likes
Arxiv
1w
59
Image Credit: Arxiv
MedSegFactory: Text-Guided Generation of Medical Image-Mask Pairs
- MedSegFactory is a versatile medical synthesis framework that generates high-quality paired medical images and segmentation masks across modalities and tasks.
- It aims to serve as an unlimited data repository, supplying image-mask pairs to enhance existing segmentation tools.
- The core of MedSegFactory is a dual-stream diffusion model, where one stream synthesizes medical images and the other generates corresponding segmentation masks.
- MedSegFactory enables on-demand generation of paired medical images and segmentation masks through user-defined prompts, facilitating scalable and high-quality data generation.
Read Full Article
3 Likes
Arxiv
1w
410
Image Credit: Arxiv
UKBOB: One Billion MRI Labeled Masks for Generalizable 3D Medical Image Segmentation
- Researchers have created the UK Biobank Organs and Bones (UKBOB), the largest labeled dataset of body organs.
- The dataset consists of 51,761 MRI 3D samples and over 1.37 billion 2D segmentation masks of 72 organs.
- The labels were generated using automatic labeling, followed by a cleaning pipeline and manual annotation for validation.
- The dataset has been used to train the Swin-BOB model for 3D medical image segmentation, achieving state-of-the-art results.
Read Full Article
24 Likes
For uninterrupted reading, download the app