A naukri.com initiative
ML News
Arxiv
1w
381
Image Credit: Arxiv
Stack Trace Deduplication: Faster, More Accurately, and in More Realistic Scenarios
- In large-scale software systems, stack traces are used to identify errors when there are no bug reports available.
- Existing deep learning-based approaches for stack trace deduplication have not been evaluated in real-life workflows.
- This work introduces a novel model and an industry-based dataset called SlowOps.
- The evaluation shows that the proposed model outperforms others in accuracy and speed.
Read Full Article
22 Likes
Arxiv
1w
243
Image Credit: Arxiv
Answer Set Networks: Casting Answer Set Programming into Deep Learning
- Answer Set Networks (ASN) is proposed as a scalable approach to ASP-based Deep Probabilistic Logic Programming (DPPL).
- ASNs leverage Graph Neural Networks (GNN) and GPU's batching and parallelization capabilities to efficiently solve encoded problems.
- Experimental evaluations demonstrate that ASNs outperform CPU-bound NeSy systems on multiple tasks.
- ASNs also contribute to the finetuning of Large Language Models (LLM) with DPPLs and the encoding of public aviation laws for drone navigation.
Read Full Article
14 Likes
Arxiv
1w
142
Image Credit: Arxiv
Surrogate-assisted multi-objective design of complex multibody systems
- The optimization of large-scale multibody systems with multiple conflicting criteria is a challenging task.
- Surrogate models, constructed from a small but informative number of expensive model evaluations, are commonly used to approximate the Pareto set of optimal compromises.
- A back-and-forth approach between surrogate modeling and multi-objective optimization is presented to improve solution quality.
- Different strategies regarding multi-objective optimization, sampling, and surrogate modeling are compared to determine the most efficient and high-quality approach.
Read Full Article
8 Likes
Arxiv
1w
410
Image Credit: Arxiv
AI-Powered Intracranial Hemorrhage Detection: A Co-Scale Convolutional Attention Model with Uncertainty-Based Fuzzy Integral Operator and Feature Screening
- Intracranial hemorrhage (ICH) refers to the leakage or accumulation of blood within the skull, which occurs due to the rupture of blood vessels in or around the brain.
- The primary aim of this study is to detect the occurrence or non-occurrence of ICH, followed by determining the type of subdural hemorrhage (SDH).
- A novel approach for ICH detection is introduced by adding two layers to the co-scale convolutional attention (CCA) classifier architecture.
- By combining features from computed tomography (CT) scan images and using a boosting neural network, the detection accuracy is improved.
Read Full Article
24 Likes
Arxiv
1w
398
Image Credit: Arxiv
From Point to probabilistic gradient boosting for claim frequency and severity prediction
- Gradient boosting algorithms have become popular in actuarial applications for their superior predictive performance.
- A comprehensive study compares various gradient boosting algorithms, including GBM, XGBoost, DART, LightGBM, CatBoost, EGBM, PGBM, XGBoostLSS, cyclic GBM, and NGBoost.
- The study assesses their performance on claim frequency and severity prediction using different datasets.
- LightGBM and XGBoostLSS are found to be computationally efficient, while EGBM achieves competitive predictive performance with interpretability.
Read Full Article
23 Likes
Arxiv
1w
305
Image Credit: Arxiv
Corn Ear Detection and Orientation Estimation Using Deep Learning
- A computer vision-based system has been developed for detecting and tracking ears of corn in an image sequence.
- The system accurately detects, tracks, and predicts the ear's orientation, providing insights into the plant's health and development.
- Using an object detector with keypoint detection, the algorithm achieved a 90 percent accuracy in ear detection.
- The system has the potential to save time and improve efficiencies in maize production.
Read Full Article
18 Likes
Arxiv
1w
180
Image Credit: Arxiv
IDOL: Instant Photorealistic 3D Human Creation from a Single Image
- Creating high-fidelity, animatable 3D full-body avatars from a single image is challenging.
- A dataset called HuGe100K with 100K diverse, photorealistic sets of human images is introduced.
- A feed-forward transformer model is developed to predict a 3D human Gaussian representation from a given image.
- The model can reconstruct photorealistic humans at 1K resolution instantly and supports various applications.
Read Full Article
10 Likes
Arxiv
1w
175
Image Credit: Arxiv
Knowledge Injection via Prompt Distillation
- Large language models (LLMs) often need to incorporate new knowledge not present in their pre-training data.
- Retrieval-augmented generation (RAG) is the industry standard for knowledge injection, but fine-tuning has not achieved comparable success.
- A new fine-tuning technique called prompt distillation is proposed to learn new knowledge and match the performance of RAG.
- Prompt distillation involves generating question-answer pairs about the new knowledge and training a student model to mimic the output distributions of a teacher model that receives the new knowledge in its prompt.
Read Full Article
10 Likes
Arxiv
1w
318
Image Credit: Arxiv
Stitch Contrast and Segment_Learning a Human Action Segmentation Model Using Trimmed Skeleton Videos
- Existing skeleton-based human action classification models rely on well-trimmed action-specific skeleton videos for both training and testing.
- A novel framework for skeleton-based action segmentation is proposed, utilizing short trimmed skeleton videos for training and longer un-trimmed videos for testing.
- The framework includes three steps: Stitch, Contrast, and Segment, which involves stitching skeleton videos, learning contrastive representations, and performing action segmentation.
- Experiments are conducted to evaluate the effectiveness of the proposed method in real-world skeleton-based human action segmentation.
Read Full Article
19 Likes
Arxiv
1w
356
Image Credit: Arxiv
DisCo: Graph-Based Disentangled Contrastive Learning for Cold-Start Cross-Domain Recommendation
- Recommender systems often face the challenge of the user cold-start problem.
- Cross-domain recommendation (CDR) is a solution to improve prediction performance in one domain using user interactions from another.
- The DisCo framework proposes a graph-based disentangled contrastive learning approach to capture user intent and avoid negative transfer.
- Experimental results demonstrate that DisCo outperforms existing baselines on benchmark CDR datasets.
Read Full Article
21 Likes
Arxiv
1w
104
Image Credit: Arxiv
Stable-V2A: Synthesis of Synchronized Sound Effects with Temporal and Semantic Controls
- Stable-V2A is a two-stage model that allows sound designers to synthesize synchronized sound effects with temporal and semantic controls.
- The first stage, RMS-Mapper, estimates an envelope representative of the audio characteristics associated with the input video.
- The second stage, Stable-Foley, is a diffusion model based on Stable Audio Open that generates audio semantically and temporally aligned with the target video.
- The model is trained and tested on the Greatest Hits dataset and a new dataset called Walking The Maps, consisting of videos extracted from video games.
Read Full Article
6 Likes
Arxiv
1w
398
Image Credit: Arxiv
DCTdiff: Intriguing Properties of Image Generative Modeling in the DCT Space
- This paper introduces DCTdiff, an end-to-end diffusion generative paradigm that models images in the DCT space.
- DCTdiff outperforms pixel-based diffusion models in terms of generative quality and training efficiency.
- It can scale up to high-resolution generation without using the latent diffusion paradigm.
- The paper also explores intriguing properties of DCT image modeling, bridging the gap between diffusion and autoregressive models.
Read Full Article
23 Likes
Arxiv
1w
113
Image Credit: Arxiv
MultiverSeg: Scalable Interactive Segmentation of Biomedical Imaging Datasets with In-Context Guidance
- MultiverSeg is a system that enables rapid segmentation of biomedical imaging datasets.
- It allows practitioners to segment a new dataset without requiring access to existing labeled data from that task or domain.
- The model takes user interactions such as clicks, bounding boxes, or scribbles as input and predicts a segmentation.
- MultiverSeg reduces the number of interactions required to segment each new image as the context set of labeled images grows.
Read Full Article
6 Likes
Arxiv
1w
159
Image Credit: Arxiv
AceMath: Advancing Frontier Math Reasoning with Post-Training and Reward Modeling
- AceMath is a suite of frontier math models designed to solve complex math problems.
- The models have been fine-tuned using a supervised fine-tuning process and targeted fine-tuning for the math domain.
- AceMath-72B-Instruct outperforms other math models, including Qwen2.5-Math-72B-Instruct, GPT-4o, and Claude-3.5 Sonnet.
- AceMath-72B-RM is a math-specialized reward model that consistently outperforms state-of-the-art reward models.
Read Full Article
9 Likes
Arxiv
1w
121
Image Credit: Arxiv
Tests for model misspecification in simulation-based inference: from local distortions to global model checks
- This study presents a simulation-based framework for model misspecification analysis in the context of simulation-based inference (SBI) techniques for Bayesian parameter estimation.
- The framework includes distortion-driven model misspecification tests and connections to classical techniques such as anomaly detection, model validation, and goodness-of-fit residual analysis.
- An efficient self-calibrating training algorithm is introduced to improve the performance of the framework.
- The framework is demonstrated in various scenarios and applied to real gravitational wave data (GW150914).
Read Full Article
7 Likes
For uninterrupted reading, download the app