Kenichi Tanaka

Harshana Habaragamuwa, Yu Oishi, Kenichi Tanaka

Agricultural image recognition tasks are becoming increasingly dependent on deep learning (DL); however, despite the excellent performance of DL, it is difficult to comprehend the type of logic or features of the input image it uses during decision making. Knowing the logic or features is highly crucial for result verification, algorithm improvement, training data improvement, and knowledge extraction. However, the explanations from the current heatmap-based algorithms are insufficient for the abovementioned requirements. To address this, this paper details the development of a classification and explanation method based on a variational autoencoder (VAE) architecture, which can visualize the variations of the most important features by visualizing the generated images that correspond to the variations of those features. Using the PlantVillage dataset, an acceptable level of explainability was achieved without sacrificing the classification accuracy. The proposed method can also be extended to other crops as well as other image classification tasks. Further, application systems using this method for disease identification tasks, such as the identification of potato blackleg disease, potato virus Y, and other image classification tasks, are currently being developed.

Kohei Numata, Kenichi Tanaka

In the field of chemistry, there have been many attempts to predict the properties of unknown compounds from statistical models constructed using machine learning. In an area where many known compounds are present (the interpolation area), an accurate model can be constructed. In contrast, data in areas where there are no known compounds (the extrapolation area) are generally difficult to predict. However, in the development of new materials, it is desirable to search this extrapolation area and discover compounds with unprecedented physical properties. In this paper, we propose Stochastic Threshold Model Trees (STMT), an extrapolation method that reflects the trend of the data, while maintaining the accuracy of conventional interpolation methods. The behavior of STMT is confirmed through experiments using both artificial and real data. In the case of the real data, although there is no significant overall improvement in accuracy, there is one compound for which the prediction accuracy is notably improved, suggesting that STMT reflects the data trends in the extrapolation area. We believe that the proposed method will contribute to more efficient searches in situations such as new material development.