Deep-learning models have become pervasive tools in science and engineering. However, their energy requirements now increasingly limit their scalability. Deep-learning accelerators aim to perform deep learning energy-efficiently, usually targeting the inference phase and often by exploiting physical substrates beyond conventional electronics. Approaches so far have been unable to apply the backpropagation algorithm to train unconventional novel hardware in situ. The advantages of backpropagation have made it the de facto training method for large-scale neural networks, so this deficiency constitutes a major impediment. Here we introduce a hybrid in situ-in silico algorithm, called physics-aware training, that applies backpropagation to train controllable physical systems. Just as deep learning realizes computations with deep neural networks made from layers of mathematical functions, our approach allows us to train deep physical neural networks made from layers of controllable physical systems, even when the physical layers lack any mathematical isomorphism to conventional artificial neural network layers. To demonstrate the universality of our approach, we train diverse physical neural networks based on optics, mechanics and electronics to experimentally perform audio and image classification tasks. Physics-aware training combines the scalability of backpropagation with the automatic mitigation of imperfections and noise achievable with in situ algorithms. Physical neural networks have the potential to perform machine learning faster and more energy-efficiently than conventional electronic processors and, more broadly, can endow physical systems with automatically designed physical functionalities, for example, for robotics, materials and smart sensors.
Download full-text PDF |
Source |
---|---|
http://www.ncbi.nlm.nih.gov/pmc/articles/PMC8791835 | PMC |
http://dx.doi.org/10.1038/s41586-021-04223-6 | DOI Listing |
Adv Sci (Weinh)
January 2025
College of Physics Science & Technology, School of Life Sciences, Institute of Life Science and Green Development, Key Laboratory of Brain-Like Neuromorphic Devices and Systems of Hebei Province, Hebei University, Baoding, 071002, China.
Hardware system customized toward the demands of graph neural network learning would promote efficiency and strong temporal processing for graph-structured data. However, most amorphous/polycrystalline oxides-based memristors commonly have unstable conductance regulation due to random growth of conductive filaments. And graph neural networks based on robust and epitaxial film memristors can especially improve energy efficiency due to their high endurance and ultra-low power consumption.
View Article and Find Full Text PDFSci Rep
January 2025
Cardiovascular Research Center, Rajaie Cardiovascular, Medical, and Research Center, University of Medical Sciences, Tehran, Iran.
Assessing myocardial viability is crucial for managing ischemic heart disease. While late gadolinium enhancement (LGE) cardiovascular magnetic resonance (CMR) is the gold standard for viability evaluation, it has limitations, including contraindications in patients with renal dysfunction and lengthy scan times. This study investigates the potential of non-contrast CMR techniques-feature tracking strain analysis and T1/T2 mapping-combined with machine learning (ML) models, as an alternative to LGE-CMR for myocardial viability assessment.
View Article and Find Full Text PDFSci Rep
January 2025
North Carolina School of Science and Mathematics, Durham, NC, 27705, USA.
Mobile Ad Hoc Networks (MANETs) are increasingly replacing conventional communication systems due to their decentralized and dynamic nature. However, their wireless architecture makes them highly vulnerable to flooding attacks, which can disrupt communication, deplete energy resources, and degrade network performance. This study presents a novel hybrid deep learning approach integrating Convolutional Neural Networks (CNN) with Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) architectures to effectively detect and mitigate flooding attacks in MANETs.
View Article and Find Full Text PDFSci Rep
January 2025
College of Computer and Information Engineering, Nanjing Tech University, Nanjing, 211800, China.
Graph data is essential for modeling complex relationships among entities. Graph Neural Networks (GNNs) have demonstrated effectiveness in processing low-order undirected graph data; however, in complex directed graphs, relationships between nodes extend beyond first-order connections and encompass higher-order relationships. Additionally, the asymmetry introduced by edge directionality further complicates node interactions, presenting greater challenges for extracting node information.
View Article and Find Full Text PDFSci Rep
January 2025
College of Computer and Data Science, Minjiang University, Fuzhou, 350018, China.
This study presents a novel approach to identifying meters and their pointers in modern industrial scenarios using deep learning. We developed a neural network model that can detect gauges and one or more of their pointers on low-quality images. We use an encoder network, jump connections, and a modified Convolutional Block Attention Module (CBAM) to detect gauge panels and pointer keypoints in images.
View Article and Find Full Text PDFEnter search terms and have AI summaries delivered each week - change queries or unsubscribe any time!