Materiality and risk in the age of pervasive AI sensors

Nature Machine Intelligence, Год журнала: 2025, Номер unknown

Опубликована: Март 20, 2025

Язык: Английский

---

Optimising TinyML with quantization and distillation of transformer and mamba models for indoor localisation on edge devices

Scientific Reports, Год журнала: 2025, Номер 15(1)

Опубликована: Март 24, 2025

Abstract This paper proposes small and efficient machine learning models (TinyML) for resource-constrained edge devices, specifically on-device indoor localisation. Typical approaches localisation rely on centralised remote processing of data transmitted from lower powered devices such as wearables. However, there are several benefits moving this to the device itself, including increased battery life, enhanced privacy, reduced latency lowered operational costs, all which key common applications health monitoring. The work focuses model compression techniques, quantization knowledge distillation, significantly reduce size while maintaining high predictive performance. We base our a large state-of-the-art transformer-based seek deploy it within low-power MCUs. also propose state-space-based architecture using Mamba more compact alternative transformer. Our results show that quantized transformer performs well 64 KB RAM constraint, achieving an effective balance between precision. Additionally, has strong performance under even tighter constraints, 32 RAM, without need compression, making viable option resource-limited environments. demonstrate that, through framework, is feasible advanced onto MCUs with restricted memory limitations. application these TinyML in healthcare potential revolutionize patient monitoring by providing accurate, real-time location minimising power consumption, increasing improving reducing infrastructure costs.

Язык: Английский

---

Deep Learning for Non-Invasive Blood Pressure Monitoring: Model Performance and Quantization Trade-Offs

Electronics, Год журнала: 2025, Номер 14(7), С. 1300 - 1300

Опубликована: Март 26, 2025

The development of non-invasive blood pressure monitoring systems remains a critical challenge, particularly in resource-constrained settings. This study proposes an efficient deep learning framework integrating Edge Artificial Intelligence for continuous estimation using photoplethysmography (PPG) signals. We evaluate three architectures: residual-enhanced convolutional neural network, transformer-based model, and attentive BPNet. Using the MIMIC-IV waveform database, we implement signal processing pipeline with adaptive filtering, statistical normalization, peak-to-peak alignment. Experiments assess varying temporal windows (10 s, 20 30 s) to optimize predictive accuracy computational efficiency. Attentive BPNet achieves best performance, systolic (SBP) yielding mean absolute error (MAE) 6.36 mmHg, diastolic (DBP) MAE 4.09 arterial (MBP) 4.56 mmHg. Post-training quantization reduces model size by 90.71% (to 0.13 MB), enabling deployment on devices. These findings demonstrate feasibility deploying learning-based edge proposed provides scalable computationally solution, offering real-time, accessible that could enhance hypertension management healthcare resource utilization.

Язык: Английский

---

Navigating the Challenges and Opportunities of Tiny Deep Learning and Tiny Machine Learning in Lung Cancer Identification

Al-Nahrain Journal for Engineering Sciences, Год журнала: 2025, Номер 28(1), С. 97 - 120

Опубликована: Апрель 7, 2025

Lung cancer is the most common dangerous disease that, if treated late, can lead to death. It more likely be successfully discovered at an early stage before it worsens. Distinguishing size, shape, and location of lymphatic nodes identify spread around these nodes. Thus, identifying lung remarkably helpful for doctors. diagnosed by expert doctors; however, their limited experience may misdiagnosis cause medical issues in patients. In line computer-assisted systems, many methods strategies used predict malignancy level that plays a significant role provide precise abnormality detection. this paper, use modern learning machine-based approaches was explored. More than 70 state-of-the-art articles (from 2019 2024) were extensively explored highlight different machine deep (DL) techniques models detection, classification, prediction cancerous tumors. The efficient model Tiny DL must built assist physicians who are working rural centers swift rapid diagnosis cancer. combination lightweight Convolutional Neural Networks resources could produce portable with low computational cost has ability substitute skill doctors needed urgent cases.

Язык: Английский

---

An IoT-Based Multimodal Wearable Framework for Real-Time Epileptic Seizures Detection Using TinyML

Lecture notes on data engineering and communications technologies, Год журнала: 2025, Номер unknown, С. 70 - 80

Опубликована: Янв. 1, 2025

Язык: Английский

---