reinforcement learning

First, I would like to introduce "Reinforcement learning with neurosymbolic AI" proposed by the IBM Research team. In this paper, we propose a reinforcement learning method called Logical Optimal Action (LOA) that utilizes Logical Neural Networks (LNN). We conclude that we can learn well.

LNN is a neuro-symbolic AI method, and was recently proposed by the IBM Research team in "Logical Neural Networks" as one of deep learning that can guarantee explainability.

LNN's revolutionary technique is to learn a formula to use in a network of models to derive inference results. A general neural network assigns weights to the features of the input data that indicate how important those features are to producing results, and learns these weights so that the prediction results are the best with a specified formula. . The logical formulas (AND, OR, NOT, etc.) used in calculation formulas are usually predefined, but LNN can learn the best logical formulas in the model. In this paper, in a text-based coin-earning game, LOA, which utilizes LNN for model reinforcement learning, LSTM-DQN++, a deep learning model for action selection, and NLM, a conventional neuro-symbolic reinforcement learning method, are presented. We compared -DQN and concluded that LOA is superior in terms of model accuracy and learning efficiency. In addition, LOA learns rules using logical symbols, and proposes that explanationable reinforcement learning can be realized in that it is possible to check the learning content.

Since AI algorithms are black Box, their explainability has been questioned for some time. The model proposed in this paper, which allows confirmation of learning content, can be expected to improve the reliability of the answers and inferences derived by AI.

Source: 34th Annual Conference of the Japanese Society for Artificial Intelligence Lecture number [3Yin2-56]
"Reinforcement Learning with Neurosymbolic AI"
Caption: Figure 1: Schematic diagram of the proposed method
https://www.jstage.jst.go.jp/article/pjsai/JSAI2022/0/JSAI2022_3Yin256/_pdf/-char/ja

dialogue system

Next, I will introduce a paper on a dialogue system that utilizes the Transformer model.

It is remarkable that communication support devices have spread in recent years due to advances in technologies such as speech recognition and language processing. Therefore, NTT Communication Science Laboratories and Tokyo Denki University researched "Construction of multimodal dialogue strategy for android robots in symbiotic society" to organize technical issues of android robots (AR) as human communication partners. , conducted an experiment using AR. This paper utilizes the Transformer-based large-scale dialogue model and BERT proposed in "Empirical Analysis of Training Strategies of Transformer-based Japanese Chit-chat Systems" published by the same author, NTT Communication Science Laboratories. and implemented a dialogue system in AR using a multimodal dialogue strategy. (See Figure 3) A schematic diagram of the interaction scenario is shown in Figure 4.

Source: The 34th Annual Conference of the Japanese Society for Artificial Intelligence Lecture number [2N5-OS-7a-02]
"Construction of Multimodal Dialogue Strategy for Android Robots in Symbiotic Society"
Caption: Figure 1: Dialogue between Android I and a customer
Figure 2: Schematic diagram of dialogue scenario
https://www.jstage.jst.go.jp/article/pjsai/JSAI2022/0/JSAI2022_2N5OS7a02/_pdf/-char/en

First, in this paper, I pointed out two issues that are rarely dealt with in ordinary robot dialogues. One is AR as a multimodal actuator and the other is AR as an asymmetric communication partner. The former is that it is necessary to implement movements such as facial expressions and voices that humans are unconsciously doing in AR, and the latter is that humans will recognize AR that simulates humans. The point is that it is necessary to perform optimal action control for AR on the premise of "difference".

As a result of verification by implementing an AR dialogue system, it was found that in the AR dialogue system, dialogue postures, nods, and facial expressions are friendly to fostering a sense of intimacy, and that communication is possible without compromising human satisfaction. showed. On the other hand, unnatural movement patterns that are not observed with human opponents are observed when AR is used as opponent, and it was concluded that optimal action control of AR assuming this is an issue for future study.

In recent years, social implementation of AR is expected as a means to solve the labor shortage due to the declining birthrate and aging population. From the experiments in this paper, we concluded that it is necessary to proceed with further verification experiments in order to acquire natural motions for the social implementation of AR. I was.

synthetic data

Finally, I would like to introduce the paper "Generation of editable medical images" by the National Cancer Center Research Institute and the RIKEN Center for Advanced Intelligence Project.

This paper focuses on the properties of generative adversarial networks (GANs) that can generate synthetic data, and proposes an approach to synthetic data generation for medical applications. GAN models have an inherent tendency to over-fit to the most frequent features in the dataset, and there is concern that synthetic data may be biased due to the small variations in case images in the medical field. I was.

Therefore, in this paper, we focus on the existence of specialized knowledge when judging diseases in clinical medicine. I suggested to generate as The generation of the segmentation map uses self-supervised learning and is editable, allowing physicians to edit directly during the data generation process, resulting in a generation algorithm that can complement their expertise.

This paper will make it possible to build a dataset that fully covers the characteristics of specific diseases in the medical field, and is expected to solve the problem of the small number of training data in medical images.

Source: 34th Annual Conference of the Japanese Society for Artificial Intelligence Lecture number [2K5-OS-1a-01]
"Editable medical image generation"
Caption: Fig.1 Neural network architecture overview.
https://www.jstage.jst.go.jp/article/pjsai/JSAI2022/0/JSAI2022_2K5OS1a01/_pdf/-char/ja

Related article

* Tech Blog AI Women's Club *
[Sports x AI] 5 examples of AI utilization in the sports scene

* Tech Blog AI Women's Club *
[Smart Building x AI] 3 Case Studies of Using AI to Address Social Issues in Smart Buildings

* Tech Blog AI Women's Club *
3 ways to improve business processes with backend operations and AI