I want to put cutting-edge AI technology into practical use in business
I would like to read some of the JSAI2020 papers
I want to introduce AI into the system
Time needed to finish reading this article
5 minutes
Introduction
Hello, I'm Makky from Macnica AI Women's Club! Lately, perhaps because I've been conscious of a "new lifestyle," I've been able to naturally maintain social distancing. The impact of COVID-19 is causing major changes not only in Japan but also in countries around the world, and it has also brought about changes such as the format of each AI academic conference becoming "online". This time, we have introducedCVPR2019orAAAI2019Similarly, we would like to introduce some of the papers accepted at JSAI2020!
Held online this year: JSAI2020
The National Conference of the Japanese Society for Artificial Intelligence (JSAI) is one of the largest AI societies in Japan. JSAI was held in Niigata in 2019, but this year it was held online due to the influence of the COVID-19. XAI (Explainable AI: explainability of deep learning), which I wrote about in my blog when I participated in JSAI2019, continues to be a trend in AI this year, and many papers on XAI were accepted. Another thing that caught my attention when I was checking the papers was that more and more papers written based on actual AI cases mentioned the systemization and operation phases.
Therefore, this time, I would like to pick up and introduce two papers related to the systemization of XAI and AI.
Paper 1: Neural generator for sparse local linear models
First, I will introduce the method proposed in "Neural generator for sparse local linear models".
In this paper ・High prediction performance of deep neural network (DNN) ・Ease of interpretation of linear models In order to realize the advantages of DNN and linear models, we are proposing a "neural generator for sparse local linear models (NGSLL)".
Configuring NGSLLs
NGSLL consists ofWGN (Weight Generator Network), which generates dense weights for each case in DNN, and K-HGM (K-Hot Gate Module), which selects only important features, that is, generates sparse weights. ). NGSLL also maintains the high prediction performance of DNNs with this approach, and achieves interpretability of prediction results because the sparse weights generated aid in the interpretation of predictions.
Source: “NEURAL GENERATORS OF SPARSE LOCAL LINEAR MODELS FOR ACHIEVING BOTH ACCURACY AND INTERPRETABILITY” Caption: Figure 1: An overview of the NGSLL for binary classification and scalar-value regression. https://arxiv.org/pdf/2003.06441.pdf
Experimental results using MNIST
The paper shows experimental results using MNIST that NGSLL achieves high prediction performance and that the generated weights help the prediction interpretation. K in the figure below indicates the number of important features selected by K-HGM. "NGSLL (w/o K-HGM)" is NGSLL without K-HGM and is an example of using dense weights.
Source: “NEURAL GENERATORS OF SPARSE LOCAL LINEAR MODELS FOR ACHIEVING BOTH ACCURACY AND NTERPRETABILITY” Caption: Figure 3: Weights visualization for binary-class MNIST dataset. https://arxiv.org/pdf/2003.06441.pdf
From the MNIST experimental results, we can see that NGSLL, which generated sparse weights with K-HGM,accurately assigns weights only to regions of black pixels. The sparse weights also make it easier to understand which regions of the image are important for prediction. The paper also includes experimental results comparing NGSLL with other interpretable methods, so be sure to read the paper for more details!
Paper 2: Consideration of Project Risk in System Development Implementing Artificial Intelligence
Even if you have not actually developed an AI system, you can imagine that there are more points to consider in system development that incorporates AI compared to conventional development projects. Even in the "AI Product Quality Assurance Guidelines"issued by theAI Product Quality Assurance Consortium, it is said that it is very difficult to understand and evaluate quality, explain and manage quality because conventional software quality assurance cannot be used.
In the next paper, "Consideration of Project Risk in System Development Implementing Artificial Intelligence," the project management of AI-introduced system development is discussed from three perspectives.
1.development framework 2.Uncertainty risk treatment 3.Miscommunication I also experienced AI introduction system development, but I feel that considerations related to project management from these three perspectives can also be applied to QA4AI (Quality Assurance for AI). .
development framework
Conventional software development followed development processes such as waterfall and agile. What is the development process like when incorporating the new technology of AI?
Source: 34th Annual Conference of Japanese Society for Artificial Intelligence Session ID: 4O3-GS-13-05 ”Consideration of Project Risk in System Development Implementing Artificial Intelligence” Caption: Fig.3 V-shaped model https://doi.org/10.11517/pjsai.JSAI2020.0_4O3GS1305
If AI PoC is not performed, the accuracy target should be set as shown in the figure above and agile development should be carried out. However, in actual development, there is a good chance that the accuracy target will not be achieved due to factors such as lack of data. The point here is that it is necessary to consider and consider in advance the accuracy that can be compromised.
Uncertainty risk response
AI does not always guarantee success, even in terms of accuracy. Therefore, there is no doubt that the problem of "uncertainty" will occur when introducing AI into the system. Uncertainty is difficult to understand, but let me give you an example from a paper.
Source: 34th Annual Conference of Japanese Society for Artificial Intelligence Session ID: 4O3-GS-13-05 ”Consideration of Project Risk in System Development Implementing Artificial Intelligence” Caption: Fig.5 Classification of uncertainty https://doi.org/10.11517/pjsai.JSAI2020.0_4O3GS1305
In this paper, we applied the perspective of AI-implemented systems to Courtney's classification of uncertainty. To reduce uncertainty, it is important to make appropriate decisions. The rationality of this decision depends on the knowledge of the decision maker. It is the actions of the project manager that matter. The paper is authored byDisseminate information that allows project managers to make appropriate decisions to decision makerssaid it was important.
Miscommunication
AI is a highly specialized field. As for myself, there were many parts of AI that I could understand by actually working with my hands and constructing an AI model. In other words, AI is a field where knowledge levels are likely to be biased, and as a result, miscommunication is likely to occur. As a specific example to prevent miscommunication in the AI introduction system development project, the paper states that it is necessary to clarify the communication plan and objectives, promote them, and coordinate them. In addition, since the risk of miscommunication can be reduced by using AI project management indicators, it is desirable to use QCDA as an AI project evaluation indicator, adding AI accuracy "Accuracy" to QCD.
Source: Source: 34th Annual Conference of Japanese Society for Artificial Intelligence Session ID: 4O3-GS-13-05 ”Consideration of Project Risk in System Development Implementing Artificial Intelligence” Caption: Fig.6 QCDA matrix https://doi.org/10.11517/pjsai.JSAI2020.0_4O3GS1305
The above figure is the QCDA matrix quoted from the paper, but each relationship can be expressed as follows. ・Cost and Delivery will increase if high accuracy is required. ・Cost and Delivery can be lowered by lowering Accuracy.
What we can see from this is that there are more things to consider in system development projects that implement AI (accuracy is increasing), so it is important to find a balance for the entire system. If you know which QCDA to prioritize, that element will also be a judgment factor as a success criterion for the project.
Looking back on the development so far, there are many points that need to be kept in mind in system development that incorporates AI compared to conventional software development, and it is hard to understand that conventional project management is not sufficient. increase. Of course, I think that the way to proceed will differ for each project, but I think that this paper will serve as a reference as one of the knowledge about how to proceed with AI projects and quality control in the future, so if you are interested, please check the paper. please look!
This time, we introduced two papers of JSAI2020. There are many interesting papers other than those introduced here, and I was able to gain a lot of knowledge while conducting research.
This year was also a year of unprecedented environmental changes, so I think that many new approaches will be announced that are necessary for changes in lifestyles. I want to keep abreast of trends and keep a close eye on them, but my body is my capital, so I always feel that I have to take care of my physical condition on a daily basis. It's getting colder now, so please take care of yourself.
■ Sources of AI papers featured in this article / Reference Lists
・Yuya Yoshikawa, Tomoharu Iwata, “NEURAL GENERATORS OF SPARSE LOCAL LINEAR MODELS FOR ACHIEVING BOTH ACCURACY AND INTERPRETABILITY”, https://arxiv.org/pdf/2003.06441.pdf
Macnica 's AI Research & Innovation Hub (ARIH) conducts cutting-edge AI research, investigations, and implementation evaluations to provide knowledge that combines the most appropriate AI technologies, and works to lead companies to optimal solutions to their problems. Please see below for details.
Latest Information
Case Study
Blog
Document List