AI Iteration Beyond Imagination: Let's Explore Neuromorphic Computing!

2019-12-20

Deep learning is blooming everywhere, but it may not be the ultimate solution for artificial intelligence. Both the academic and industrial sectors are contemplating the next development path for AI, and neuromorphic computing has quietly become one of the "seed players" that attracts attention. Recently, the "Beijing High-End Forum and 2019 Future Chip Forum" was jointly organized by the Beijing Future Chip Technology High-End Innovation Center and the Tsinghua University Institute of Microelectronics. At this forum, neuromorphic computing became a hot topic discussed by authoritative experts. Neuromorphic chips are a type of architecture in artificial intelligence chips. They are designed to simulate the human brain and have greater advantages in power consumption and learning ability compared to traditional chips. What is neuromorphic computing? Neuromorphic computing, also known as neuromorphic computing, is not only a new hot topic in academic conferences but also being explored by the industry. The research on neuromorphic chips is based on the combination of microelectronics technology and new neural morphological devices. It aims to break through the traditional computing architecture, achieve deep integration of storage and computation, greatly improve computing performance, increase integration, and reduce energy consumption. By drawing on the basic rules of information processing in the brain, it fundamentally changes the existing computing systems and achieves significant improvements in computing energy consumption, computing power, and computing efficiency at multiple levels, such as hardware implementation and software algorithms. If the current wave of AI development is a turbulent wave, then neuromorphic computing is the undercurrent beneath it. Although it may not attract much attention now, it has the potential to change the trend of AI development in the future. One reason is that although deep learning has made great breakthroughs and has been widely applied in fields such as speech recognition, image recognition, and natural language understanding, it requires a large amount of computing power and has high power consumption. A professor from Peking University's Department of Computer Science and Technology gave a vivid example: intelligent drones that apply deep learning technology are already very agile in the market, but in terms of intelligence, they are far behind a fly or dragonfly, despite their much larger size and power consumption. Neuromorphic computing attempts to simulate the structure and information processing of biological neural networks. One of its attempts at the software level is the pulse neural network. Currently, deep learning is generally implemented through convolutional neural networks (CNN) or recurrent neural networks (RNN). CNN and RNN are both artificial neural networks, but the artificial neurons used in them still rely on models from the 1940s. Although the designed artificial neural networks are becoming larger and more complex, the fundamental neural models have not been significantly improved. On the other hand, in deep learning artificial neural networks, the connections between neurons are called weights, which are crucial elements of artificial neural networks. In pulse neural networks, however, the connections between neurons are neural pulses, and information processing and expression are achieved through neural pulses. Just like in our brains, there are a large number of neural pulses transmitting and circulating. Neuromorphic intelligence has good application prospects, and many excellent domestic companies have achieved remarkable results in the development, mass production, and AI architecture of neuromorphic chips. The information and communication technology revolution has gone through the processes of electronics, digitalization, networking, and intelligence. Nowadays, computers are becoming smaller, and the end of Moore's Law is approaching, with the physical limitations becoming increasingly clear. Some people are implementing neuromorphic computing from the hardware level, such as neural morphological chips. In July 2019, Intel announced that its neuromorphic research chip, Loihi, can perform specialized tasks 1,000 times faster and 10,000 times more efficiently than a regular CPU. In the traditional von Neumann architecture, information processing and storage are separate, while in the human brain, they are integrated. Therefore, chips with integrated storage and computation have become the fundamental solution to avoid constant data transfer within the chip and can improve the chip's energy efficiency. Currently, neuromorphic computing is still in the exploratory stage and lacks typical successful applications. However, commercial companies have already sensed the potential, and it may not take long for related technologies to achieve large-scale applications. The development of neuromorphic computing is still relatively preliminary compared to existing mainstream AI algorithms, and the main challenge is the lack of talent. From algorithms and computing frameworks to the implementation of the entire chip, China lacks talent in all aspects and needs comprehensive and vigorous talent cultivation to support the AI industry. From the perspective of deep learning, we are much later than foreign countries in terms of starting, and from the platform perspective, we have our own platforms that can almost dominate the field of deep learning. However, in the field of neuromorphic intelligence, there is no dominant player yet. New helium neuromorphic intelligence, Fudan University's neuromorphic center, and other universities and institutions like Tsinghua University are leading the way in China's neuromorphic intelligence. At this stage, people need more efficient neural network hardware to promote the manufacturing of neuromorphic chips. Professor Anders Lansner from Stockholm University said that neuroscience can help us achieve this goal. To truly understand the operation mechanism of the human brain, we need to combine neuroscience with cognitive psychology, and for the future generation of AI, we need a deep understanding of the content and concepts behind these disciplines.

RECENT AI NEWS

RECENT AI TOOLS