AI technologies handling unstructured cognitive data (e.g voices, images, videos, and texts) are becoming a crucial source of competitiveness for all industries in both cloud and edge environments. Artificial neural networks are a powerful concept to handle such data, however, current implementations by software and general-purpose processors with GPU acceleration are power-hungry and computer-intensive so far because of the recent surge in the amount of unstructured data and the growing scale of AI models. The mainstream of current hardware platforms for AI is based on well-matured ASIC technology and CMOS technologies supported by Moore’s law. However, with the downscaling limit of conventional technologies, traditional digital electronic computing will face performance and power efficiency challenges, because of the constraints imposed by the von-Neumann architecture. Thus, developing disruptive AI hardware platforms for these challenges is essential to realize next-generation cloud and edge computing systems.
To address the power and performance constraints, acceleration by neuromorphic devices is a promising approach. In fact, various CMOS-based neuromorphic devices have been reported so far. In recent years, motivated by the potential computational capabilities of various natural physical phenomena, unconventional computing paradigms have been actively investigated in the interdisciplinary region of computer science and natural science. For example, optical/photonics neuromorphic computing based on matured silicon photonics and optical interconnect is attracting much attention as a high-speed AI accelerator. Analog memristive devices are utilized for neuromorphic computing with on-chip learning capability. Physical reservoir computing utilizes a broader range of physics such as photonics, spintronics, material, and quantum for speed and power efficiency. The objectives of this Research Topic are to investigate the possibility of incorporating such diverse natural physical phenomena into neuromorphic computing and to go beyond the limitation of traditional CMOS-based electronics devices, which we call “physical neuromorphic computing”. In this Research Topic, to accelerate the development of physical neuromorphic computing from diverse backgrounds, the Topic Editors welcome papers to present the state-of-the-art technologies relevant to theory, algorithm, implementation, and practical industry applications.
The scope of interest in this Research Topic includes the following, but is not limited to:
1. Algorithms and device implementation of physical neuromorphic computing such as:
- electronic neuromorphic computing
- optical/photonic neuromorphic computing
- physical reservoir computing (optics/photonics, magnetics, materials, mechanics, etc.)
One can use major standard benchmark tasks such as image and voice recognition for validation and comparison purposes. We welcome articles that describe novel datasets for neuromorphic systems including MNIST, EMNIST and N-MNIST datasets.
2. Industrial applications of physical neuromorphic computing, in particular, cloud and edge computing such as:
- acceleration of large-scale AI algorithms at cloud
- Internet-of-Things, sensor data analytics
- surveillance, anomaly detection at edge
- control of autonomous vehicles, robots, and drones
- intelligent networking systems, in-network computing
Since our Research Topic covers wide application areas as above, one can use open-data sources, for example, text corpora, network traffic data, machine-to-machine, and IoT transaction data, etc.
3. Neuroscience-specific applications of physical neuromorphic computing
- Simulations of neural networks
- Analysis of cognitive data including neural imaging data
- Use of data sets for brain models and visualization
AI technologies handling unstructured cognitive data (e.g voices, images, videos, and texts) are becoming a crucial source of competitiveness for all industries in both cloud and edge environments. Artificial neural networks are a powerful concept to handle such data, however, current implementations by software and general-purpose processors with GPU acceleration are power-hungry and computer-intensive so far because of the recent surge in the amount of unstructured data and the growing scale of AI models. The mainstream of current hardware platforms for AI is based on well-matured ASIC technology and CMOS technologies supported by Moore’s law. However, with the downscaling limit of conventional technologies, traditional digital electronic computing will face performance and power efficiency challenges, because of the constraints imposed by the von-Neumann architecture. Thus, developing disruptive AI hardware platforms for these challenges is essential to realize next-generation cloud and edge computing systems.
To address the power and performance constraints, acceleration by neuromorphic devices is a promising approach. In fact, various CMOS-based neuromorphic devices have been reported so far. In recent years, motivated by the potential computational capabilities of various natural physical phenomena, unconventional computing paradigms have been actively investigated in the interdisciplinary region of computer science and natural science. For example, optical/photonics neuromorphic computing based on matured silicon photonics and optical interconnect is attracting much attention as a high-speed AI accelerator. Analog memristive devices are utilized for neuromorphic computing with on-chip learning capability. Physical reservoir computing utilizes a broader range of physics such as photonics, spintronics, material, and quantum for speed and power efficiency. The objectives of this Research Topic are to investigate the possibility of incorporating such diverse natural physical phenomena into neuromorphic computing and to go beyond the limitation of traditional CMOS-based electronics devices, which we call “physical neuromorphic computing”. In this Research Topic, to accelerate the development of physical neuromorphic computing from diverse backgrounds, the Topic Editors welcome papers to present the state-of-the-art technologies relevant to theory, algorithm, implementation, and practical industry applications.
The scope of interest in this Research Topic includes the following, but is not limited to:
1. Algorithms and device implementation of physical neuromorphic computing such as:
- electronic neuromorphic computing
- optical/photonic neuromorphic computing
- physical reservoir computing (optics/photonics, magnetics, materials, mechanics, etc.)
One can use major standard benchmark tasks such as image and voice recognition for validation and comparison purposes. We welcome articles that describe novel datasets for neuromorphic systems including MNIST, EMNIST and N-MNIST datasets.
2. Industrial applications of physical neuromorphic computing, in particular, cloud and edge computing such as:
- acceleration of large-scale AI algorithms at cloud
- Internet-of-Things, sensor data analytics
- surveillance, anomaly detection at edge
- control of autonomous vehicles, robots, and drones
- intelligent networking systems, in-network computing
Since our Research Topic covers wide application areas as above, one can use open-data sources, for example, text corpora, network traffic data, machine-to-machine, and IoT transaction data, etc.
3. Neuroscience-specific applications of physical neuromorphic computing
- Simulations of neural networks
- Analysis of cognitive data including neural imaging data
- Use of data sets for brain models and visualization