Abstract:Neuromorphic computing takes inspiration from the brain to create energy efficient hardware for information processing, capable of highly sophisticated tasks. In this article, we make the case that building this new hardware necessitates reinventing electronics. We show that research in physics and material science will be key to create artificial nano-neurons and synapses, to connect them together in huge numbers, to organize them in complex systems, and to compute with them efficiently. We describe how some researchers choose to take inspiration from artificial intelligence to move forward in this direction, whereas others prefer taking inspiration from neuroscience, and we highlight recent striking results obtained with these two approaches. Finally, we discuss the challenges and perspectives in neuromorphic physics, which include developing the algorithms and the hardware hand in hand, making significant advances with small toy systems, as well as building large scale networks.
What problem does this paper attempt to address?
This paper aims to explore how to draw inspiration from the brain to create more efficient information - processing hardware, especially for the field of neuromorphic computing. The paper points out that the existing electronic technology is not sufficient to meet this need, because traditional computer architectures are not only slow but also energy - consuming when processing complex cognitive tasks. In contrast, the brain can complete complex tasks with extremely low energy consumption. Therefore, research in physics and materials science is crucial for developing new artificial nano - neurons and synapses, which can be integrated on a large scale and organized and computed in an efficient manner.
Specifically, the paper discusses the following points:
1. **Why draw inspiration from the brain**:
- The biological brain, due to its survival needs, can perform extremely complex tasks and compute with a very low energy budget.
- The human brain can complete tasks such as classification, prediction, and creation under approximately 20 watts of power, while modern supercomputers need to consume 1,000 kilowatt - hours of energy when training the most advanced natural language processing models, which is equivalent to the energy required for the human brain to process all tasks throughout its entire life cycle.
2. **The brain's algorithms and physical implementation**:
- The brain is a dynamic, reconfigurable complex system that can learn through experience.
- Physicists, through statistical physics, nonlinear dynamics, and complex - system theory, have revealed how neural mechanisms support learning.
- The physical implementation of the brain is based on biological entities, such as synapses and neurons, rather than memory blocks and transistors in digital circuits. It utilizes the randomness of cells to work at extremely low energy, using binary and analog encoding, as well as asynchronous communication methods.
3. **Why physics and materials science are crucial for neuromorphic computing**:
- Current electronic technology cannot effectively solve the problem of the separation of memory and computing, resulting in the "von Neumann bottleneck" and limiting the performance of artificial intelligence algorithms.
- The goal of neuromorphic computing is to imitate the topological structure of the brain, construct circuits composed of physical neurons and synapses, achieve in - situ non - volatile storage, thereby greatly reducing the need for data transmission, and improving speed and energy efficiency.
- It is necessary to develop devices that can imitate the characteristics of neurons and synapses at the nano - scale. These devices should have the characteristics of low energy consumption, high durability, easy addressing, providing signal gain, and storage functions.
4. **Two main research methods**:
- Map existing neural network algorithms to dedicated physical systems to improve the energy efficiency when running these algorithms.
- Draw inspiration from neuroscience to add new features and dynamics to artificial neural networks to achieve more complex computations.
5. **The application of materials and physical phenomena**:
- The paper details the applications of various materials and physical phenomena in neuromorphic computing, including memristors, phase - change memories, two - dimensional materials, organic electrochemical devices, superconducting Josephson junctions, etc.
- These materials and devices can imitate the functions of neurons and synapses, achieving efficient multiply - accumulate operations, low - power - consumption signal processing, and learning capabilities.
In conclusion, this paper emphasizes the crucial role of physics and materials science in neuromorphic computing and proposes the directions and challenges for future research.