Marking a significant breakthrough for China in brain-like computing and perception.
Recently, Tsinghua University has achieved an important breakthrough in the field of brain-like visual perception chips:
The team led by Professor Shi Luping from the Brain-like Computing Research Center at the Department of Precision Instruments at Tsinghua University, proposed a new paradigm of brain-like visual perception based on visual primitives with complementary dual pathways, and developed the world's first brain-like complementary vision chip "Tianmou Core". The paper based on this research "A Vision Chip with Complementary Pathways for Open-world Sensing" was featured on the cover of the May 30th issue of the journal "Nature".
This is the second time the team has been featured on the cover of "Nature" after the heterogeneous integration of brain-like computing "Tianji Core", marking a significant breakthrough for China in the direction of brain-like computing and perception.
With the rapid development of artificial intelligence, unmanned systems such as autonomous driving and embodied intelligence are continuously being promoted and applied in real society, leading a new round of technological revolution and industrial transformation. In these intelligent systems, visual perception, as the core way to obtain information, plays an extremely important role. However, in complex and unpredictable environments, achieving efficient, accurate and robust visual perception (that is, the system's ability to survive under abnormal and dangerous conditions) remains a formidable challenge.
Advertisement
In the open world, intelligent systems not only need to process a large amount of data but also need to deal with various extreme events, such as sudden dangers in driving, intense light changes at the entrance of tunnels, and strong flash interference at night. Traditional visual perception chips are often faced with distortion, failure or high latency when dealing with these scenarios due to the limitations of "power consumption wall" and "bandwidth wall", which seriously affects the stability and safety of the system.To overcome these challenges, the Center for Brain-inspired Computing Research in the Department of Precision Instruments at Tsinghua University has focused on brain-inspired visual perception chip technology, proposing a new paradigm of brain-inspired visual perception based on visual primitives and complementary dual pathways. This paradigm draws on the basic principles of the human visual system, decomposing visual information in the open world into representations based on visual primitives, and mimicking the characteristics of the human visual system by organically combining these primitives to form two visually perceptive pathways that are complementary in advantages and complete in information.
Based on this new paradigm, the team further developed the world's first brain-inspired complementary vision chip, "Tianmou Core," achieving high-speed visual information acquisition at 10,000 frames per second with high precision of 10 bits and high dynamic range of 130 dB, at the cost of extremely low bandwidth (reduced by 90%) and power consumption.
It not only breaks through the performance bottleneck of traditional visual perception paradigms but also can efficiently deal with various extreme scenarios to ensure the stability and safety of the system. Based on "Tianmou Core," the team has independently developed high-performance software and algorithms, and conducted performance verification on an open environment vehicle platform. Under various extreme scenarios, the system has achieved low-latency, high-performance real-time perception and inference, demonstrating its great potential for application in the field of intelligent unmanned systems.
The successful development of "Tianmou Core" is undoubtedly a major breakthrough in the field of intelligent perception chips. It not only provides strong technical support for the development of the intelligent revolution but also opens up new paths for important applications such as autonomous driving and embodied intelligence. Combined with the team's accumulated technical experience in brain-inspired computing chips "Tianji Core," brain-inspired software toolchains, and brain-inspired robots, the addition of "Tianmou Core" will further improve the brain-inspired intelligence ecosystem and powerfully promote the development of artificial general intelligence.
Professor Shi Luping and Professor Zhao Rong from the Department of Precision Instruments at Tsinghua University are the corresponding authors of the paper, and Dr. Yang Zhe Yu (currently the R&D manager of Beijing Lingxi Technology Co., Ltd.), doctoral student Wang Taoyi from the Department of Precision Instruments, and Lin Yihan are the co-first authors of the paper. Tsinghua University is the first unit of the paper, and the cooperative units include Beijing Lingxi Technology Co., Ltd.
This research was supported by the Ministry of Science and Technology's Innovation 2030 "Brain Science and Brain-inspired Research" major project and the National Natural Science Foundation of China, and also supported by the Tsinghua University/IDG-McGovern Institute for Brain Research.
Tsinghua University published two papers in Nature on the same day.In addition, another latest scientific research achievement from Tsinghua University was published in "Nature" at the same time.
The team led by Lu-Ming Duan published a research paper titled "A site-resolved two-dimensional quantum simulator with hundreds of trapped ions" online in Nature. The research reported the stable trapping of 512 ions in a two-dimensional Wigner crystal and the sideband cooling of their lateral motion.
Quantum computing and quantum simulation have entered an era of hundreds of qubits, with complex computational tasks beyond the capabilities of current classical computers. To further expand the application of noise intermediate-scale quantum (NISQ) devices on practical problems and classical puzzles, quantum simulation of many-body dynamics and NISQ algorithms such as quantum annealing and variational quantum algorithms have aroused widespread research interest. In addition to a large number of qubits, a key requirement for these applications is the ability to read out the state of individual qubits in a single shot, allowing, for example, to measure the spatial correlation of qubits under quantum dynamics or to evaluate the many-body target function in optimization problems.
As one of the leading quantum computing platforms, ion traps have demonstrated quantum simulation with up to 61 qubits in one-dimensional (1D) Paul traps and have single detection. To further expand the number of qubits, a feasible approach is to capture ions in two-dimensional (2D) crystals. So far, 2D crystals containing up to 150 ions under Doppler cooling, about 50 ions under two-tone laser cooling, and about 100 ions under electromagnetically induced transparency (EIT) cooling have been reported, and global quantum manipulation and individual detection have also been demonstrated for up to 10 ions in two dimensions.
It is worth noting that two-dimensional micro-trap arrays have also been implemented on a small scale, with larger spacing and weak coupling strength, and two-dimensional junctions of quantum charge coupling device architecture have also been realized as a feasible method for expanding systems. In contrast, two-dimensional ion crystals in Penning traps are native, and quantum simulation of about 200 ions has been achieved. However, due to the rapid rotation of the ion crystal in the Penning trap, the detection of individual qubit states remains an experimental challenge. Although spatial and temporal resolution imaging techniques have been used to calculate the number of ions, the observation results of quantum simulation in the Penning trap are still limited to global.
The report presents the stable trapping of a two-dimensional ion crystal of 512 171Yb+ ions, with the associated lateral mode (perpendicular to the ion plane) EIT and sideband cooling to less than one phonon per mode. The researchers demonstrated quantum simulation of remote quantum Ising models with tunable coupling strength and modes, with or without frustration, using 300 ions. By using the position resolution in a single measurement, rich spatial correlation patterns were observed in the ground state prepared quasi-adiabatically, which allowed the verification of quantum simulation results by comparing the measured two-spin correlations with the calculated collective phonon modes and classical simulated annealing. The quenching dynamics of the Ising model in the transverse field were further explored to demonstrate quantum sampling tasks. The research paves the way for simulating classical intractable quantum dynamics and running noisy intermediate-scale quantum algorithms with two-dimensional ion trap quantum simulators.
