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

Large language models (LLMs) have shown remarkable capabilities in a wide range of natural language processing tasks. Recently, there has been growing interest in the use of reinforcement learning (RL) and planning to improve the performance and usability of LLMs. In this talk, I will provide a brief overview of the state-of-the-art in this area and discuss some of the challenges and open questions. In particular, I will cover challenges and best practices when training LLM-based reward models for RL as well as how to run reward-model based RLHF. I will also discuss the nascent area of planning algorithms for language models, and particularly fruitful directions for future work.

Abstract

Quantum computing’s rapid advancement threatens current cryptography, which relies on complex problems solvable by quantum computers. This necessitates the development of “quantum-safe” technologies for future network protection. Today’s data is also at risk from “harvest now, decrypt later” attacks, where encrypted messages are stored for decryption once quantum processors are available. This is especially concerning for long-term valuable information like financial records and medical data.

Quantum cryptography offers a set of protocols with provable security. Specifically, Quantum Key Distribution (QKD) allows secret keys to be shared on optical networks, making it immune to attacks by powerful processors, including quantum computers.

This talk will review quantum cryptography’s recent progress, focusing on integrating quantum communications into conventional networks. It will feature real-world examples of deployment in metropolitan networks and use-cases in finance, healthcare, and government sectors.

Abstract

Many practical problems require choosing from a very large number of possible solutions, but current computing is extremely limited in the scale of the problems it can handle. While quantum computing offers the promise of solving large-scale combinatorial optimization problems quickly, there is a need for quantum-inspired solutions that could be deployed in the near-term. In this talk we will review the state of the art in hardware designs for simulated annealing, a meta-heuristic solver for combinatorial optimization problems and describe an architecture based on stochastic computing for the efficient hardware implementation of annealing. We will also discuss applications of simulated annealing in emerging communications systems.

Abstract

Large language models have achieved impressive performance across the NLP task spectrum and even appear to have superhuman conversational capabilities. However, they suffer from hallucinations, lack of transparency, and limited ability to improve over time without needing full and rather expensive retraining. Modular dialogue systems, on the other hand, offer an interpretable underlying state and action and a set-up for long-term reward optimisation via reinforcement learning. In this talk, I’ll explain the steps we take towards building a continually learning task-oriented dialogue system consisting of a dynamic policy model, a data-driven user simulator, and a challenging environment to study the ability of the system to learn in a world that is continuously changing.

Abstract

Large language models (LLMs), a technology that underpins ChatGPT, have emerged as standout advancements in the AI research field in recent years. Indeed, its success captures the attention of many companies and research institutes eager to develop LLMs. It is well-known that building an LLM necessitates extensive computational infrastructure and substantial budgets (financial investments), but it also requires considerable expertise and tremendous effort from the developers involved. This is because learning LLMs tends to be unstable for several reasons. Moreover, most LLM developers are new to the task and lack the knowledge shared in the community to overcome the issue. In this talk, I intend to delve into the common challenges associated with the instability encountered during the training phase of LLMs. First, I will briefly describe the recent achievements of LLMs and then introduce some typical observations in training LLMs from scratch. Finally, I will show various strategies devised to address or at least alleviate the instability problems of LLMs. Through this talk, I would like to share the difficulty of building a new LLM.

Abstract

Living systems are cell-based and reproduce themselves through a network of chemical transformations (metabolism), whereby the offspring have the same proliferation abilities as the parent structure (recursive reproduction). The chemical network is maintained by numerous proteins that are generated based on sequence information encoded in DNA. A traditional approach to understanding living systems is to reconstruct the metabolic pathways that synthesize lipids using proteins expressed by DNA inside a vesicular compartment so that vesicle reproduction occurs. However, since this approach utilizes the heart of a complex living system (i.e., the central dogma of molecular biology), shedding light on the road from non-living forms of matter to living systems is still very challenging. In this study, we present a synthetic minimal cell system approach, where vesicles reproduce themselves based on instructions encoded in information molecules using non-biological substances. Although our minimal cells still have significant gaps from living systems, we have succeeded in reproducing the essence of important life systems, such as the linkage between replicating information molecules and the self-production of vesicles. We believe this approach sheds light on the pathway from non-living systems to living systems.

Abstract

We have proposed a coherently linked optical and wireless transmission system toward a large-capacity radio access network (RAN) beyond 5G or6G. In this talk, we present our recent work on broadband, long-reach mobile fronthaul with digital coherent optical transmission technologies and efforts toward optical-wireless converged coherent access network.