東北大学人工知能エレクトロニクス卓越大学院プログラム

Abstract

Technology helps visually impaired people become more independent than ever before. Voice-based Web access allows them to get all the information they need. Smartphones offer new experiences such as accessing information, communicating with others via SNS, and controlling home appliances in a mobile environment. AI is now offering new experiences for the visually impaired to recognize and understand their surrounding world. For example, computer vision technology starts describing scenes and people’s behavior around them, and recognizing their personal items, such as T-shirts, mugs and so on. The visually impaired users can expect to walk and travel around the world independently, naturally, and safely.
During her talk, she will first introduce how technology has changed her life. She will then talk about her latest research on the real-world accessibility, followed by the new challenges caused by the pandemic. Finally, she will discuss how we can accelerate the implementation of new technologies into society to achieve truly inclusive society.

Abstract

EdgeML combines the power of machine (deep) learning and edge (IoT) devices. Owing to its capability of solving difficult problems in sensor nodes and other resource constrained devices, EdgeML has seen adoption in variety of application domains like smart manufacturing, remote monitoring, smart homes etc However, being deployed on edge devices exposes machine/deep learning algorithms to range of new attacks specially physical attacks.

In this talk, we demonstrate practical physical attacks on EdgeML. First, we show how side-channel attacks can be used to reverse engineer architectures and parameters of deep learning models. These models are often proprietary with commercial value and contain information on sensitive training data. The feasibility of these attacks are shown both on standalone microcontrollers as well as commercial ML accelerators. Further, we demonstrate practical and low-cost cold boot based model recovery attacks on Intel Neural Compute Sticks 2 (NCS2) to recover the model architecture and weights, loaded from the Raspberry Pi with high accuracy. The proposed attack remains unaffected by the model encryption features of the NCS2 framework.

Abstract

Cognitive dynamic system (CDS) is inspired by human brain and it is built on the principles of perception, memory, perception-action cycle, attention and language. In this talk, the CDS concept is applied to a fiber optic system. Typical systems that use CDS such as radio and radar are linear in nature. However, fiber optic system is nonlinear and the noise is non-Gaussian. The CDS concept is extended to non-Gaussian and nonlinear environment (NGNLE). The CDS has three main blocks – (i) preceptor which perceives the environment (i.e. fiber optic channel) and develops its statistical model, (ii) executive which takes action on the fiber optic system such as change the data rate/modulation format, and (iii) main feedback channel which sends internal reward from preceptor to executive. It is shown that the quality of service (QoS) can be significantly improved in a fiber optic system upgraded with CDS.

Abstract

Thanks to the rapid advance of artificial intelligence technologies, there is a explosion of interest in research and development of social conversation machines (aka. chatting machines) that can freely and appropriately converse to humans in a natural language. Empathy, the capability to understand and resonate with emotions and experience of other people, plays an important role in human conversations. It directly affects the way people interact with each other. Thus, a human-like chatting machine must be empowered with the necessary emotional intelligence in order to offer the continuous and enjoyable conversation experience. Not to mention how significant it is in the applications like accompanying lonely elderly, providing consulting services and appeasing angry customers. There is a practical need to develop novel emotion-enhanced social conversation models that enable machines to converse to humans properly in both content and emotion. This is exactly what we are targeted to. While previous works on empathetic response generation have focused merely on the emotional state of the user, leading to weak empathy and inane content, we incorporate emotion-intervened causes and effects into the model for empathetic responses with engaging content.

Abstract

In the super-smart society that Japan aims at, Artificial Intelligence technology and data analysis are essential, but they have certain limits. The versatility and rigorousness of Mathematics have received attention as a way to break the limits and change our future business and life dramatically. Mathematics can make great contributions to innovation. However, unfortunately, the impact of Mathematics has not been so big, especially in Japan, because Mathematics has not actively collaborated with other fields, and other fields have not been fully utilizing Mathematics. In order to overcome the situation, we need new doctoral talents in Mathematics, so to speak, “mathematical modeling talents”. This is the main reason why we started the “Graduate Program of Mathematics for Innovation” in Kyushu University. The five skills that this program aims to cultivate are called “Mathematics Five Forces”. Having excellent mathematical and statistical skills is the basis and represents the soil, and by utilizing the mathematical modeling power and the power to co-create across organizations and fields, it fertilizes the lush leaves to create innovation and blooms flowers in each field, in turn cultivating “Math-for-Innovation Professionals”.

Abstract

There has been tremendous progress in developing AI that understands objects, scenes, and more from their images. We may be able to see its applications to solving practical problems soon. The key to success is the interplay between visual and linguistic information. For instance, what the AI understands can only be revealed through natural language communication with humans (and other AI agents). I will discuss how far we (i.e., the research community) have come and what remains to be done while introducing the research activity of our research group.

Abstract

Healthcare monitoring systems based on mid-infrared spectroscopy are introduced. Those systems utilize compact quantum cascade lasers, hollow optical fibers transmitting mid-infrared light, and a semiconductor detector working at room temperature. The system configurations and experimental results on measurement of blood glucose level and cholesterol amount are reported.

Abstract

Nonvolatile spintronic devices have potential advantages such as fast read/write, and high endurance together with back-end-of-the-line compatibility, which offers the possibility of constructing not only stand-alone RAMs and embedded RAMs that can be used in conventional VLSI circuits and systems, but also standby-power-free and high-performance VLSI processors, where a new logic-circuit style, called "Nonvolatile (NV) Logic-in-Memory (LIM) Architecture" that merges nonvolatile storage elements into CMOS logic-gate circuits is used [1],[2]. The advantages of employing spintronic devices, especially magnetic tunnel junction (MTJ) devices combined with CMOS circuits are discussed and the current status of the MTJ-based VLSI computing paradigm is presented along with its prospects and remaining challenges.
In my presentation, some AI-hardware examples based on MTJ-based LIM architecture are presented and their suitability for Internet-of-Things (IoT) applications are discussed [3]-[4].
[1] T. Hanyu, et al., Proc. IEEE, l. vol.104, no.10, 1844/1863, Oct. 2016.
[2] M. Natsui, et al., IEEE JSSC, vol.54, no.11, 2991/3004, Aug. 2019.
[3] M. Natsui, et al., JJAP, STAP article, vol.59, 050602(7pages), April 2020.
[4] D. Suzuki, et al., NOLTA, IEICE, vol.E12-N, no.4, 695/710, Oct. 2021.