Hoon-Hee KimDue to the development of machine learning such as a deep neural network, Artificial Intelligence(AI) has been used in many areas. Modern AI technology accurately solves problems such as classification, regression, and prediction, but there is a lack of skill to explain the process of AI decision in terms of human understanding, it is called a black-box AI. The black-box AI, in which humans cannot understand the decision process, is difficult to use in high- risk areas such as important social and legal decisions, medical diagnosis, and financial predictions [1]. Although there are highly explainable machine learning methods such as a decision-tree, these machine learning methods tend to has a low performance and are not suitable for solving a complex problem [2]. In this study, I suggest a novel explainable AI method which has a high performance based on an integrated model of Reservoir Computing and Autoencoder. Reservoir Computing, a recurrent neural network consists of three layers: inputs, reservoir, and readouts can train nonlinear dynamics using linear learning methods [3]. Recently, a study was published in which neural networks induced actual physical laws using Variational Autoencoder which can extract interpretable features of the learning data [4]. In the integrated model, the features of the training data were learned by the autoencoder structure and linear learning rule of reservoir computing. Therefore, these features could be represented as a linear formula form that a human can simply understand. To validate the integrated model, I tested the model to predict trends of the S&P500 index. The model showed more than 80% accuracy and reported that which features were most important to the prediction in terms of weighted linear formula.AcknowledgmentsThis study was supported by the National Research Foundation of Korea [NRF-2019R1A6A3A01096892].References1\. Rudin, C. Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead. Nature Machine Intelligence. 2019, 1(5), 206-215.2\. Defense Advanced Research Projects Agency. Broad Agency Announcement, Explainable Artificial Intelligence (XAI), DARPA-BAA-16-53 (DARPA, 2016); https://www.darpa.mil/attachments/DARPA-BAA-16-53.pdf3\. Jaeger, H. and H. Haas. Harnessing nonlinearity: predicting chaotic systems and saving energy in wireless communication. Science. 2004, 304(5667), 78-80.4\. Iten Raban, et al. Discovering Physical Concepts with Neural Networks. Physical Review Letters. 2020, 124, 010508-1-6
