本研究では、深層学習により産業連関表の投入係数行列を推計する新たな手法を提示する。この研究では都道府県のデータを用いた深層学習を試みるが、そのデータが小さいために過学習を引き起こす可能性が高い。そこで、複数の都道府県を一つにまとめた仮想的な地域のデータを生成するデータ拡張により、データサイズを増加させ過学習の回避を試みる。山口県および岐阜県郡上市について投入係数行列を予測した結果、深層学習による推計方法はRAS法より安定した精度で投入係数を予測できうることを示した。

The Tiger Pufferfish (Takifugu rubripes) is a staple in Japanese cuisine, with over ten species of the Takifugu genus found in the surrounding seas. Given that certain parts of the pufferfish are toxic, they are predominantly prepared by trained professionals. Furthermore, species within the Takifugu genus are susceptible to hybridization, leading to an increase in hybrid numbers. However, identifying these hybrids is a challenging and time-consuming task, even for experts. To address this, we developed a transfer learning model using the pretrained VGG16 model to differentiate between pufferfish species. The VGG16 model, commonly used in image recognition, is built on convolutional neural networks. We also implemented Gradient-weighted Class Activation Mapping (Grad-CAM) for visual interpretation of the model. Grad-CAM generates a heat map that highlights the areas focused on by the AI model in the image, allowing us to identify factors contributing to misjudgment and make further improvements. We used seven species from the Takifugu genus (excluding hybrids), and approximately 15 colored images of each species were prepared for machine learning. The results showed that our model was able to distinguish between pufferfish species with relatively high accuracy, although some misclassification occurred among species with similar body patterns. The Grad-CAM results revealed that the model was able to distinguish body patterns, but some misclassifications occurred due to gravel and background objects being recognized as patterns.