High-Mix Low-Volume Oriented Edge AI with Low-Cost and Its Evaluation

Smart cities are being promoted due to the declining population. In smart cities, AI using sensing information such as camera image information provides various services. However, for financial reasons, it is important to realize various types of AI with limited use (high-mix low-volume) at low-cost as AI in smart cities. In terms of the cost of AI generation, it is necessary to consider the design cost and the training cost. In addition, high resistance to adversarial examples (adversarial robustness) is also important for practical applications of AI. Adversarial training is reported to be effective in achieving high adversarial robustness. Therefore, this study proposes a low-cost high-mix low-volume edge AI as a smart city-oriented AI. The proposed edge AI is based on VAT, which combines semi-supervised learning and AT to reduce learning cost. The proposed edge AI consists of a binary neural network (BNN) and uses high-level synthesis for its implementation to reduce design cost. This reduces design and training costs and creates a secure edge AI that can handle a small volume and variety of products.