Breaking the data barrier: a review of deep learning techniques for democratizing AI with small datasets

Artificial Intelligence Review, Journal Year: 2024, Volume and Issue: 57(9)

Published: Aug. 2, 2024

Abstract Justifiably, while big data is the primary interest of research and public discourse, it essential to acknowledge that small remains prevalent. The same technological societal forces generate datasets also produce a more significant number datasets. Contrary notion inherently superior, real-world constraints such as budget limitations increased analytical complexity present critical challenges. Quality versus quantity trade-offs necessitate strategic decision-making, where often leads quicker, accurate, cost-effective insights. Concentrating AI research, particularly in deep learning (DL), on exacerbates inequality, tech giants Meta, Amazon, Apple, Netflix Google (MAANG) can easily lead due their access vast datasets, creating barrier for mid-sized enterprises lack similar access. This article addresses this imbalance by exploring DL techniques optimized offering comprehensive review historic state-of-the-art models developed specifically study aims highlight feasibility benefits these approaches, promoting inclusive equitable landscape. Through PRISMA-based literature search, 175+ relevant articles are identified subsequently analysed based various attributes, publisher, country, utilization dataset technique, size, performance. delves into current highlights open problems, recommendations future investigations. Additionally, importance developing effectively utilize domains acquisition difficult expensive.

Language: Английский

---

Interpretable vertical federated learning with privacy-preserving multi-source data integration for prognostic prediction

Engineering Applications of Artificial Intelligence, Journal Year: 2025, Volume and Issue: 148, P. 110408 - 110408

Published: March 9, 2025

Language: Английский

---

Structural digital Twin for damage detection of CFRP composites using meta transfer Learning-based approach

Expert Systems with Applications, Journal Year: 2024, Volume and Issue: unknown, P. 125458 - 125458

Published: Oct. 1, 2024

Language: Английский

---

MIF: Multi-source information fusion for few-shot classification with CLIP

Pattern Recognition Letters, Journal Year: 2025, Volume and Issue: unknown

Published: April 1, 2025

Language: Английский

---

Inverse design of non-parametric acoustic metamaterials via transfer-learned dual variational autoencoder with latent space-based data augmentation

Engineering Applications of Artificial Intelligence, Journal Year: 2025, Volume and Issue: 151, P. 110735 - 110735

Published: April 8, 2025

Language: Английский

---

Transfer Learning With GANs and Meta-Learning

Advances in computational intelligence and robotics book series, Journal Year: 2025, Volume and Issue: unknown, P. 233 - 258

Published: Feb. 21, 2025

This chapter explains how combined Transfer Learning, GANs, and Meta-Learning the artificial intelligence fields where it can help solve challenging problems. In situations information is scanty, characteristics of a model increase through Learning across contexts. GAN generator-discriminator networks generate realistic synthetic populations, whereas demands prompt adaptation to address new learning problems with little samples. By answering these questions, this will discover solves data scarcity, enhances generalization, foster advent medical imaging as well creativity. The responsible innovation subject acquisition diverse cross-disciplinary knowledge. Lastly, might develop revolutionary AI that private public efficiency improve creativity many areas.

Language: Английский

---

FishMet: A Digital Twin Framework for Appetite, Feeding Decisions and Growth in Salmonid Fish

Aquaculture Fish and Fisheries, Journal Year: 2025, Volume and Issue: 5(2)

Published: April 1, 2025

ABSTRACT Salmonids are important fish species in aquaculture countries the temperate zone. Optimisation of feeding next‐generation precision farming requires developing models for decision support and process control. Black box ML AI often very efficient but have drawbacks, such as requiring large amount training data reduced performance novel situations where no available. Thus, realistic appetite, decisions, feed intake, energetics growth is necessary. Such essential predicting performance, example, waste from uneaten faeces, growth, ‘what if’ scenario testing. We built a conceptual model based on review major neurophysiological mechanisms feedback loops controlling appetite food intake fish. Building this, we developed FishMet model: new extensible stochastic simulation framework that represents basic energy budget salmonid The advance, while bioenergetic part follows established theory. supported by server‐based components open API assimilation on‐demand execution allows to use digital twin. demonstrate relatively good prediction stomach gut digesta transit rainbow trout Oncorhynchus mykiss . twin also demonstrated efficiency pilot scale experiment Atlantic salmon Salmo salar discuss concept directions further development an applied predictive tool.

Language: Английский

---

Metric-Based Meta-Learning Approach for Few-Shot Classification of Brain Tumors Using Magnetic Resonance Images

Electronics, Journal Year: 2025, Volume and Issue: 14(9), P. 1863 - 1863

Published: May 2, 2025

Brain tumor prediction from magnetic resonance images is an important problem, but it difficult due to the complexity of brain structure and variability in appearance. There have been various ML DL-based approaches, limitations current models are a lack adaptability new tasks need for extensive training on large datasets. To address these issues, novel meta-learning approach has proposed, enabling rapid adaptation with limited data. This paper presents method that integrates vision transformer metric-based model, few-shot learning enhance classification performance. The proposed begins preprocessing MRI images, followed by feature extraction using transformer. A Siamese network enhances model’s learning, quick unseen data improving robustness. Furthermore, applying strategy performance when there comparison other developed reveals consistently performs better. It also compared previously approaches same datasets evaluation metrics including accuracy, precision, specificity, recall, F1-score. results demonstrate efficacy our methodology classification, which significant implications enhancing diagnostic accuracy patient outcomes.

Language: Английский

---

IFTrans: A few-shot classification method for consumer fraud detection in social sensing

Knowledge-Based Systems, Journal Year: 2025, Volume and Issue: unknown, P. 113608 - 113608

Published: May 1, 2025

Language: Английский

---

Model and Method for Providing Resilience to Resource-Constrained AI-System

Sensors, Journal Year: 2024, Volume and Issue: 24(18), P. 5951 - 5951

Published: Sept. 13, 2024

Artificial intelligence technologies are becoming increasingly prevalent in resource-constrained, safety-critical embedded systems. Numerous methods exist to enhance the resilience of AI systems against disruptive influences. However, when resources limited, ensuring cost-effective becomes crucial. A promising approach for reducing resource consumption during test-time involves applying concepts and dynamic neural networks. Nevertheless, networks various disturbances remains underexplored. This paper proposes a model architecture training method that integrate with focus on resilience. Compared conventional methods, proposed yields 24% increase convolutional 19.7% visual transformers under fault injections. Additionally, it results 16.9% network ResNet-110 21.6% transformer DeiT-S adversarial attacks, while saving more than 30% computational resources. Meta-training improves task changes by an average 22%, achieving same level savings.

Language: Английский

---