(Invited) Random Network Structure of Materials Developing Next Generation AI Devices for Autonomous Robotics

Abstract

1. IntroductionIn recent years, the superior computational power of deep learning based on software has been widely recognized, and the practical applications of artificial intelligence are rapidly expanding. On the other hand, the hardware for replacing to such artificial intelligence (AI) algorithms is facing the physical limits of scaling in silicon CMOS technology, and performance improvement is expected to hit the ceiling. For this reason, there is a growing interest in hardware technologies that physically implement artificial neural networks (ANNs), neuromorphic or brainmorphic information processing systems, and the applications (hereafter referred as AI systems in this paper), as well as new materials and devices. A critical difference between the presently required device functionality and that in conventional computational systems is the use of dynamics. By cleverly using nanomaterials' nonlinearity and network structure, devices that spontaneously generate pulses, noise, and other physical phenomena are expected to be realized to utilize for the AI hardware. These devices will enable drastically lower power consumption and higher integration of AI systems. In the learning process of ANNs, it is necessary to constantly change and store the weights of the weighted sum (sum-of-products) part. In our research center, we have been working on materials that can complement CMOS for AI systems by using molecules and nanocarbon materials, and further, we are trying to apply them to autonomous AI robots. This paper introduces these nanomaterials and networks’ formation as devices, the key points of the devices’ functionalization, application to robots, and other recent research results.2. Formation and Control of NanomaterialsOrganic and inorganic nanosized molecules can be synthesized to form structures by the current development of synthesis technology; a great variety of molecules can be realized and controlled for a high potential to use as neuromorphic devices. In order to elicit electronic functionalities, it is necessary to transfer electrons to and from the molecules, which is a redox reaction from a chemical point of view. Stable Carbon-based materials that have a close Fermi level to the molecules may be more suitable than metallic electrodes. We have connected nanoparticles of organic molecules to single-walled carbon nanotubes (SWNTs) to control the electrical properties of the nanoscale [1].3. Results obtained from Random NetworksWe formed a random network consisting of SWNTs and POMs and found that it generates neuron-like pulses when a high voltage is applied (Fig. 1, left) [2]. This enabled us to control noise and pulses generation. Such dynamic signal generation is attributed to the multiple discharges of POM molecules. Reservoir computing simulations also showed that this system could be used for time series memory [2].In this study, we applied a sinusoidal signal to an SWNT/Por-POM random network using porphyrin-gradient sandwich POM (Por-POM), which is more redox-sensitive, and the Lissajous plot of the output signal from one of the multiple electrodes plotted against the input signal showed nonlinear switching behavior. A memory state was generated due to the reversible redox of Por-POM, and the input signal was highly interacted with the reservoir system. The interaction can be also checked by the changing of the Lissajous plot shape from linear to elliptical, which indicate the characteristics obtained from current and past inputs through the echo state network. The dynamics are also confirmed by the higher harmonic generation (HHG) from the FFT analysis, and NARMA-2 task which considers states up to a two-step delay. The NARMA-2 task successfully generated signals that predictably follow the target waveform using a linear regression model. These results indicate that the nonlinearity of SWNT/Por-POM and memory integrity are important factors for reservoir operation. We also used Ag/Ag2S core-shell nanoparticles (NP) as a reservoir device (Fig.1, right) to obtain similar results with SWNT/Por-POM.As mentioned above, large power consumption during the sum-of-products operation in the weighted addition part has been a problem in AI systems. While, by using reservoir computing devices consist of random networks of SWNT/POM, core-shell nanoparticles, nanowires, etc., it is expected that the sum-of-products operation can reduce power consumption. Voice recognition and robot-hand grabbing stuff recognition [3] was conducted by the reservoir devices. Details will be presented at the presentation.Refs. [1] H. Tanaka et al., Adv. Mat. 18, 1411 (2006). [2] H. Tanaka et al., Nat. Commun. 9, 2693 (2018). The article was selected as the most read 50 articles published in Nat. Commun. in 2018 (Physics). [3] Kyutech won the RoboCup world series of Domestic Standard Platform League by TOYOTA HSR in 2017 and 2018. The same robot is used for the demonstration. Figure 1