A naukri.com initiative
ML News
Arxiv
2d
90
Image Credit: Arxiv
Scaling Laws for Uncertainty in Deep Learning
- Deep learning has revealed scaling laws that predict model performance based on dataset and model sizes.
- Researchers are exploring whether similar scaling laws govern predictive uncertainties in deep learning.
- In identifiable parametric models, scaling laws for uncertainty can be derived by treating model parameters in a Bayesian way.
- Guarantees on uncertainty contraction rates do not hold in over-parameterized models.
- Empirical evidence shows scaling laws for predictive uncertainty with respect to dataset and model sizes.
- Experiments on vision and language tasks confirm scaling laws for predictive uncertainty using Bayesian inference and ensemble methods.
- This research challenges skepticism towards Bayesian approaches in deep learning.
- Having a large amount of data is not sufficient to eliminate epistemic uncertainty.
Read Full Article
5 Likes
Arxiv
2d
0
Image Credit: Arxiv
HopaDIFF: Holistic-Partial Aware Fourier Conditioned Diffusion for Referring Human Action Segmentation in Multi-Person Scenarios
- Action segmentation is a core challenge in high-level video understanding, focusing on partitioning videos and assigning predefined action labels.
- Existing methods mainly address single-person activities, leaving out multi-person scenarios.
- A new dataset, RHAS133, is introduced for Referring Human Action Segmentation in multi-person settings, comprising 133 movies with annotations for 137 actions and textual descriptions.
- Benchmarking existing methods on the RHAS133 dataset shows limited performance in aggregating visual cues for target individuals.
- To improve action segmentation in multi-person scenarios, a new framework called HopaDIFF is proposed.
- HopaDIFF leverages a holistic-partial aware Fourier-conditioned diffusion approach and a novel cross-input gate attentional xLSTM for enhanced long-range reasoning.
- The framework introduces a Fourier condition to gain more control and improve action segmentation generation.
- HopaDIFF achieves state-of-the-art results on the RHAS133 dataset across various evaluation scenarios.
- The code for HopaDIFF is available at https://github.com/KPeng9510/HopaDIFF.git.
Read Full Article
Like
Arxiv
2d
375
Image Credit: Arxiv
DipLLM: Fine-Tuning LLM for Strategic Decision-making in Diplomacy
- DipLLM is a fine-tuned Large Language Model (LLM) designed for strategic decision-making in Diplomacy, a complex multiplayer game that combines cooperation and competition.
- Traditional methods for AI in Diplomacy rely on equilibrium search, requiring extensive game data and computational resources.
- LLMs offer an alternative by leveraging pre-trained knowledge for strong performance with limited fine-tuning.
- However, applying LLMs to Diplomacy is challenging due to the game's complexity and strategic interactions among players.
- DipLLM simplifies the task by using an autoregressive factorization framework to break down multi-unit action assignment into unit-level decisions.
- The model fine-tunes by learning equilibrium policies and outperforms the Cicero model with only 1.5% of the training data.
- This research demonstrates the potential of fine-tuned LLMs for complex strategic decision-making in multiplayer games.
Read Full Article
22 Likes
Arxiv
2d
359
Image Credit: Arxiv
Intent Factored Generation: Unleashing the Diversity in Your Language Model
- Obtaining multiple diverse samples from Large Language Models for a prompt is a challenge.
- Current methods focus on token-level diversity, leading to repetitive responses.
- Intent Factored Generation (IFG) proposed to address diversity and engagement issues.
- IFG involves sampling a semantic intent first and then generating a response based on the intent and prompt.
- Higher temperature used for intent step to promote diversity, lower temperature for final generation for coherence.
- Prompting the model to state its intent before generating enhances reasoning tasks.
- IFG shows effectiveness in improving pass@k and Reinforcement Learning on math and code tasks.
- IFG combined with Direct Preference Optimization enhances conversational diversity without loss in reward.
- IFG maintains diversity and quality in general language modeling using reader comments and news articles dataset.
- IFG is a simple method to boost diversity in Large Language Models while preserving performance.
- The method is easy to integrate into various algorithms for improved performance across applications.
Read Full Article
21 Likes
Arxiv
2d
240
Image Credit: Arxiv
CINeMA: Conditional Implicit Neural Multi-Modal Atlas for a Spatio-Temporal Representation of the Perinatal Brain
- Researchers have introduced CINeMA, a novel framework for creating high-resolution, spatio-temporal, multimodal brain atlases suitable for low-data settings.
- CINeMA operates in latent space, avoiding the need for compute-intensive image registration and reducing atlas construction times from days to minutes.
- The framework allows flexible conditioning on anatomical features like gestational age, birth age, and brain pathologies such as ventriculomegaly and agenesis of the corpus callosum.
- CINeMA supports tasks like tissue segmentation, age prediction, synthetic data creation, and anatomically informed data augmentation.
- The framework surpasses existing methods in accuracy, efficiency, and versatility, making it a valuable tool for advancing brain research.
- The code and atlases for CINeMA are available at https://github.com/m-dannecker/CINeMA.
Read Full Article
14 Likes
Arxiv
2d
296
Image Credit: Arxiv
Adding simple structure at inference improves Vision-Language Compositionality
- Dual encoder Vision-Language Models (VLM) like CLIP face challenges with compositionality affecting their retrieval performance.
- Various training methods have been suggested to enhance the vision-language compositionality of these models.
- This study focuses on adding simple structure during inference to address the compositionality issue.
- The proposed method involves dividing images into smaller crops, extracting text segments describing objects, attributes, and relations, and aligning image crops with text segments using a VLM.
- The final image-text similarity is calculated by aggregating individual similarities of matched image crops and text segments.
- The approach is evaluated on popular dual encoder VLMs across controlled and natural datasets for vision-language compositionality, showing consistent performance improvements without additional training.
- Significant enhancements are observed in attribute-object binding, particularly in the controlled dataset.
- Analysis reveals the importance of processing image crops for performance gains and highlights areas for further improvement in inference-time techniques.
Read Full Article
17 Likes
Arxiv
2d
3
Image Credit: Arxiv
Empirical and computer-aided robustness analysis of long-step and accelerated methods in smooth convex optimization
- This work examines the robustness of different first-order optimization methods in dealing with relative inexactness in gradient computations.
- Three major families of methods analyzed are constant step gradient descent, long-step methods, and accelerated methods.
- Long-step and accelerated methods are theoretically shown to be not robust to inexactness initially.
- A semi-heuristic shortening factor is introduced to enhance the theoretical guarantees of long-step and accelerated methods.
- All methods are tested on an inexact problem, showing that accelerated methods are more robust than expected and the shortening factor helps long-step methods significantly.
- The study concludes that all shortened methods appear promising, even in an inexact setting.
Read Full Article
Like
Arxiv
2d
15
Image Credit: Arxiv
Alice and the Caterpillar: A more descriptive null model for assessing data mining results
- Researchers introduce novel null models for assessing data mining results using statistical hypothesis testing.
- These null models preserve more properties of observed binary transactional and sequence datasets compared to existing models.
- The new models maintain the Bipartite Joint Degree Matrix of the dataset's corresponding bipartite (multi-)graph.
- Preserving properties like the number of caterpillars (paths of length three) is a focus of the new null models.
- The researchers developed a suite named Alice, leveraging Markov chain Monte Carlo algorithms for sampling datasets from the null models.
- Alice is based on a well-defined set of states and efficient operations for transitioning between them.
- Experimental results demonstrate that Alice mixes quickly, scales effectively, and uncovers different significant results compared to existing models.
Read Full Article
Like
Arxiv
2d
11
Image Credit: Arxiv
Learning to Optimize Package Picking for Large-Scale, Real-World Robot Induction
- Warehouse automation using machine learning models can enhance operational efficiency and reduce costs in large-scale robotic fleets.
- Current research focuses on increasing picking success rates by prioritizing high-probability picks but lacks data-driven optimization for performance at scale.
- A new ML-based framework was developed to predict transform adjustments and optimize suction cup selection for multi-suction end effectors in packages.
- The framework was tested in workcells resembling Amazon Robotics' Robot Induction fleet, leading to a 20% decrease in pick failure rates compared to heuristic methods.
Read Full Article
Like
Arxiv
2d
110
Image Credit: Arxiv
Automatic Treatment Planning using Reinforcement Learning for High-dose-rate Prostate Brachytherapy
- Researchers investigated the use of reinforcement learning (RL) for high-dose-rate (HDR) prostate brachytherapy to optimize needle placement based on patient anatomy.
- The RL agent adjusts needle positions and dwell times to maximize a reward function, with multiple rounds played until all needles are optimized.
- Data from 11 patients were included, showing RL plans had similar prostate coverage and rectum dose compared to clinical plans, but lower prostate hotspot and urethra dose.
- RL plans used, on average, two fewer needles than clinical plans, demonstrating potential for improved efficiency and plan quality.
- This study showcases the feasibility of RL in autonomously generating practical HDR prostate brachytherapy plans, offering standardized planning and enhanced patient outcomes.
Read Full Article
6 Likes
Arxiv
2d
280
Image Credit: Arxiv
CoRT: Code-integrated Reasoning within Thinking
- Large Reasoning Models (LRMs) like o1 and DeepSeek-R1 have made progress in natural language reasoning with long chain-of-thought (CoT) but struggle with complex mathematical operations.
- Code Interpreter (CI) introduces external knowledge to LRMs, but combining it directly poses challenges.
- CoRT is a post-training framework designed to teach LRMs to effectively use CI for complex mathematical operations.
- Data scarcity is addressed by synthesizing code-integrated reasoning data through Hint-Engineering, strategically inserting hints to optimize LRM-CI interaction.
- 30 high-quality samples are manually created to post-train models ranging from 1.5B to 32B parameters using supervised fine-tuning, rejection fine-tuning, and reinforcement learning.
- Hint-Engineering models show 4% and 8% absolute improvements on DeepSeek-R1-Distill-Qwen-32B and DeepSeek-R1-Distill-Qwen-1.5B respectively across five challenging mathematical reasoning datasets.
- Hint-Engineering models use about 30% fewer tokens for the 32B model and 50% fewer tokens for the 1.5B model compared to natural language models.
- Experimental results demonstrate the effectiveness of CoRT in improving LRMs' performance on mathematical reasoning tasks.
- The models and code for CoRT are available at https://github.com/ChengpengLi1003/CoRT.
Read Full Article
16 Likes
Arxiv
2d
150
Image Credit: Arxiv
A Deep Generative Model for the Simulation of Discrete Karst Networks
- The simulation of discrete karst networks is challenging due to complex physicochemical processes in geological and hydrogeological contexts.
- A novel approach using graph generative models to represent karst networks has been proposed.
- Karst networks are represented as graphs with nodes containing spatial information and edges indicating connections between nodes.
- The generative process involves utilizing graph recurrent neural networks (GraphRNN) to learn the topological distribution of karst networks.
- Denoising diffusion probabilistic models on graphs (G-DDPM) are used to learn node features like spatial coordinates.
- The approach aims to generate realistic karst networks that capture essential features of the original data.
- Real-world karst networks were used to test the approach by comparing generated subgraphs with actual subgraphs.
- Geometry and topology metrics were employed to evaluate the generated subgraphs.
- The methodology allows for stochastic simulation of discrete karst networks across various formations.
- It serves as a useful tool for studying physical processes such as flow and transport in karst environments.
Read Full Article
9 Likes
Arxiv
2d
304
Image Credit: Arxiv
Advancing Exchange Rate Forecasting: Leveraging Machine Learning and AI for Enhanced Accuracy in Global Financial Markets
- Foreign exchange rate forecasting, like USD to BDT, is critical in global financial markets impacting trade and economic stability.
- This study uses historical USD/BDT data from 2018-2023 to develop machine learning models for accurate forecasting.
- A Long Short-Term Memory (LSTM) neural network achieves 99.449% accuracy, with an RMSE of 0.9858, outperforming ARIMA.
- A Gradient Boosting Classifier (GBC) is also employed for directional prediction, revealing a 40.82% profitable trade rate.
- Historical trends analysis shows a decline in BDT/USD rates and incorporates normalized daily returns for volatility.
- Deep learning in forex forecasting offers traders and policymakers robust tools to mitigate risks in financial markets.
- Future work may involve integrating sentiment analysis and real-time economic indicators for enhanced model adaptability.
Read Full Article
18 Likes
Arxiv
2d
276
Image Credit: Arxiv
PersonaLens: A Benchmark for Personalization Evaluation in Conversational AI Assistants
- Large language models have improved conversational AI assistants, but evaluating personalization in these assistants is challenging.
- Existing personalization benchmarks do not capture the complexities of personalized task-oriented assistance.
- To address this gap, PersonaLens is introduced, a benchmark for evaluating personalization in task-oriented AI assistants.
- PersonaLens includes diverse user profiles with rich preferences and interaction histories, along with specialized LLM-based user and judge agents.
- The user agent engages in realistic task-oriented dialogues with AI assistants, while the judge agent assesses personalization, response quality, and task success.
- Extensive experiments with current LLM assistants across diverse tasks have shown significant variability in personalization capabilities.
- PersonaLens provides crucial insights for the advancement of conversational AI systems.
Read Full Article
16 Likes
Arxiv
2d
35
Image Credit: Arxiv
Kvasir-VQA-x1: A Multimodal Dataset for Medical Reasoning and Robust MedVQA in Gastrointestinal Endoscopy
- Kvasir-VQA-x1 is a new multimodal dataset designed for medical reasoning and robust MedVQA in gastrointestinal endoscopy.
- The dataset addresses the limitations of current datasets by incorporating 159,549 new question-answer pairs to test deeper clinical reasoning.
- Questions in the dataset are stratified by complexity to evaluate a model's inference capabilities more effectively.
- To prepare models for real-world clinical scenarios, visual augmentations that simulate common imaging artifacts have been included in the dataset.
- Kvasir-VQA-x1 supports two evaluation tracks: one for standard VQA performance and the other to assess model robustness against visual perturbations.
- The dataset aims to accelerate the development of more reliable and effective AI systems for clinical use by providing a challenging and clinically relevant benchmark.
- Kvasir-VQA-x1 adheres to FAIR data principles, ensuring accessibility and transparency for the wider research community.
- Code and data related to the dataset can be found on GitHub at https://github.com/Simula/Kvasir-VQA-x1
- Access to the dataset is available at https://huggingface.co/datasets/SimulaMet/Kvasir-VQA-x1
Read Full Article
2 Likes
For uninterrupted reading, download the app