A new dataset for measuring the performance of blood vessel segmentation methods under distribution shifts DOI Creative Commons
Matheus Viana da Silva, Natália de Carvalho Santos,

Julie Ouellette

et al.

PLoS ONE, Journal Year: 2025, Volume and Issue: 20(5), P. e0322048 - e0322048

Published: May 27, 2025

Creating a dataset for training supervised machine learning algorithms can be demanding task. This is especially true blood vessel segmentation since one or more specialists are usually required image annotation, and creating ground truth labels just single take up to several hours. In addition, it paramount that the annotated samples represent well different conditions might affect imaged tissues as possible changes in acquisition process. only achieved by considering typical atypical, even outlier, samples. We introduce VessMAP, an highly heterogeneous acquired carefully sampling relevant images from large non-annotated containing fluorescence microscopy images. Each of contains metadata information regarding contrast, amount noise, density, intensity variability vessels. Prototypical atypical were selected base using available information, thus defining assorted set used measuring performance on distinct each other. show datasets traditionally developing new tend have low heterogeneity. Thus, neural networks trained few four generalize all other VessMAP critical generalization capability network. For instance, with good contrast leads models poor inference quality. Interestingly, while some sets lead Dice scores 0.59, careful selection results score 0.85. development active methods selecting manual annotation analyzing robustness distribution shifts data.

Language: Английский

DepGraph: Towards Any Structural Pruning DOI
Gongfan Fang, Xinyin Ma, Mingli Song

et al.

2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Journal Year: 2023, Volume and Issue: unknown, P. 16091 - 16101

Published: June 1, 2023

Structural pruning enables model acceleration by removing structurally-grouped parameters from neural networks. However, the parameter-grouping patterns vary widely across different models, making architecture-specific pruners, which rely on manually-designed grouping schemes, non-generalizable to new architectures. In this work, we study a highly-challenging yet barely-explored task, any structural pruning, tackle general of arbitrary architecture like CNNs, RNNs, GNNs and Transformers. The most prominent obstacle towards goal lies in coupling, not only forces layers be pruned simultaneously, but also expects all removed consistently unimportant, thereby avoiding issues significant performance degradation after pruning. To address problem, propose fully automatic method, Dependency Graph (DepGraph), explicitly dependency between comprehensively group coupled for extensively evaluate our method several architectures tasks, including ResNe(X)t, DenseNet, MobileNet Vision transformer images, GAT graph, DGCNN 3D point cloud, alongside LSTM language, demonstrate that, even with simple norm-based criterion, proposed yields gratifying performances.

Language: Английский

Citations

213
Machine Learning-Guided Protein Engineering DOI Creative Commons
Petr Kouba, Pavel Kohout, Faraneh Haddadi

et al.

ACS Catalysis, Journal Year: 2023, Volume and Issue: 13(21), P. 13863 - 13895

Published: Oct. 13, 2023

Recent progress in engineering highly promising biocatalysts has increasingly involved machine learning methods. These methods leverage existing experimental and simulation data to aid the discovery annotation of enzymes, as well suggesting beneficial mutations for improving known targets. The field protein is gathering steam, driven by recent success stories notable other areas. It already encompasses ambitious tasks such understanding predicting structure function, catalytic efficiency, enantioselectivity, dynamics, stability, solubility, aggregation, more. Nonetheless, still evolving, with many challenges overcome questions address. In this Perspective, we provide an overview ongoing trends domain, highlight case studies, examine current limitations learning-based We emphasize crucial importance thorough validation emerging models before their use rational design. present our opinions on fundamental problems outline potential directions future research.

Language: Английский

Citations

98
Automated data processing and feature engineering for deep learning and big data applications: A survey DOI Creative Commons
Alhassan Mumuni, Fuseini Mumuni

Journal of Information and Intelligence, Journal Year: 2024, Volume and Issue: unknown

Published: Jan. 1, 2024

Modern approach to artificial intelligence (AI) aims design algorithms that learn directly from data. This has achieved impressive results and contributed significantly the progress of AI, particularly in sphere supervised deep learning. It also simplified machine learning systems as process is highly automated. However, not all data processing tasks conventional pipelines have been In most cases be manually collected, preprocessed further extended through augmentation before they can effective for training. Recently, special techniques automating these emerged. The automation driven by need utilize large volumes complex, heterogeneous big applications. Today, end-to-end automated based on (AutoML) are capable taking raw transforming them into useful features Big Data intermediate stages. this work, we present a thorough review approaches pipelines, including preprocessing– e.g., cleaning, labeling, missing imputation, categorical encoding–as well (including synthetic generation using generative AI methods) feature engineering–specifically, extraction, construction selection. addition specific tasks, discuss use AutoML methods tools simultaneously optimize stages pipeline.

Language: Английский

Citations

40
Slimmable Dataset Condensation DOI
Songhua Liu, Jingwen Ye,

Runpeng Yu

et al.

2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Journal Year: 2023, Volume and Issue: unknown, P. 3759 - 3768

Published: June 1, 2023

Dataset distillation, also known as dataset condensation, aims to compress a large into compact synthetic one. Existing methods perform condensation by assuming fixed storage or transmission budget. When the budget changes, however, they have repeat synthesizing process with access original datasets, which is highly cumbersome if not infeasible at all. In this paper, we explore problem of slimmable extract smaller given only previous results. We first study limitations existing algorithms on such successive compression setting and identify two key factors: (1) inconsistency neural networks over different times (2) underdetermined solution space for data. Accordingly, propose novel training objective explicitly account both factors. Moreover, datasets in our method adopt significance-aware parameterization. Theoretical derivation indicates that an upper-bounded error can be achieved discarding minor components without training. Alternatively, allowed, strategy serve strong initialization enables fast convergence. Extensive comparisons ablations demonstrate superiority proposed multiple benchmarks.

Language: Английский

Citations

34
Minimizing the Accumulated Trajectory Error to Improve Dataset Distillation DOI
Jiawei Du, Yidi Jiang, Vincent Y. F. Tan

et al.

2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Journal Year: 2023, Volume and Issue: unknown

Published: June 1, 2023

Model-based deep learning has achieved astounding successes due in part to the availability of large-scale real-world data. However, processing such massive amounts data comes at a considerable cost terms computations, storage, training and search for good neural architectures. Dataset distillation thus recently come fore. This paradigm involves distilling information from large datasets into tiny compact synthetic that latter ideally yields similar performances as former. State-of-the-art methods primarily rely on dataset by matching gradients obtained during between real these gradient-matching suffer so-called accumulated trajectory error caused discrepancy subsequent evaluation. To mitigate adverse impact this error, we propose novel approach encourages optimization algorithm seek flat trajectory. We show weights trained are robust against errors perturbations with regularization towards Our method, called Flat Trajectory Distillation (FTD), is shown boost performance up 4.7% subset images ImageNet higher resolution images. also validate effectiveness generalizability our method different resolutions demonstrate its applicability architecture search. Code available at. https://github.com/AngusDujw/FTD-distillation.

Language: Английский

Citations

31
Play it Straight: An Intelligent Data Pruning Technique for Green-AI DOI Creative Commons
Francesco Scala, Sergio Flesca, Luigi Pontieri

et al.

Lecture notes in computer science, Journal Year: 2025, Volume and Issue: unknown, P. 69 - 85

Published: Jan. 1, 2025

Language: Английский

Citations

1
Data-Centric Green Artificial Intelligence: A Survey DOI
Shirin Salehi, Anke Schmeink

IEEE Transactions on Artificial Intelligence, Journal Year: 2023, Volume and Issue: 5(5), P. 1973 - 1989

Published: Sept. 14, 2023

With the exponential growth of computational power and availability large-scale datasets in recent years, remarkable advancements have been made field artificial intelligence (AI), leading to complex models innovative applications. However, these consume a significant unprecedented amount energy, contributing greenhouse gas emissions growing carbon footprint AI industry. In response, concept green has emerged, prioritizing energy efficiency sustainability alongside accuracy related measures. To this end, data-centric approaches are very promising reduce consumption algorithms. This paper presents comprehensive overview technologies their impact on Specifically, it focuses methods that utilize training data an efficient manner improve We identified multiple approaches, such as active learning, knowledge transfer/sharing, dataset distillation, augmentation, curriculum learning can contribute development environmentally-friendly implementations machine Finally, practical applications highlighted, challenges future directions discussed.

Language: Английский

Citations

24
Deep Graph Reprogramming DOI
Yongcheng Jing,

Chongbin Yuan,

Li Ju

et al.

2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Journal Year: 2023, Volume and Issue: unknown

Published: June 1, 2023

In this paper, we explore a novel model reusing task tailored for graph neural networks (GNNs), termed as "deep reprogramming". We strive to reprogram pretrained GNN, without amending raw node features nor parameters, handle bunch of cross-level downstream tasks in various domains. To end, propose an innovative Data Reprogramming paradigm alongside Model paradigm. The former one aims address the challenge diversified feature dimensions on input side, while latter alleviates dilemma fixed per-task-per-model behavior side. For data reprogramming, specifically devise elaborated Meta-FeatPadding method deal with heterogeneous dimensions, and also develop transductive Edge-Slimming well inductive Meta-GraPadding approach diverse homogenous samples. Meanwhile, task-adaptive Reprogrammable-Aggregator, endow frozen larger expressive capacities handling cross-domain tasks. Experiments fourteen datasets across node/graph classification/regression, 3D object recognition, distributed action demonstrate that proposed methods yield gratifying results, par those by re-training from scratch.

Language: Английский

Citations

19
Knowledge fusion distillation and gradient-based data distillation for class-incremental learning DOI
Lin Xiong, Xin Guan, Hailing Xiong

et al.

Neurocomputing, Journal Year: 2025, Volume and Issue: 622, P. 129286 - 129286

Published: Jan. 5, 2025

Language: Английский

Citations

1
Leveraging synthetic data to tackle machine learning challenges in supply chains: challenges, methods, applications, and research opportunities DOI Creative Commons
Y. F. Long, Sebastian Kroeger, Michael F. Zaeh

et al.

International Journal of Production Research, Journal Year: 2025, Volume and Issue: unknown, P. 1 - 22

Published: Jan. 8, 2025

Machine learning (ML) has the potential to improve various supply chain management (SCM) tasks, namely demand forecasting, risk management, inventory production planning and control, network reconstruction, distribution logistics. However, industrial application of ML in chains faces many challenges, particularly data privacy scarcity. Synthetic data, which is artificially generated mimic real-world patterns, shown promise overcoming similar challenges fields such as healthcare finance. synthetic context remains limited. This publication aims analyze machine operations (MLOps) for tasks explain how can address these a context. Moreover, identify suitable approaches generate data. Based on analysis, research agenda proposed guideline future activities enable use chains.

Language: Английский

Citations

1