LegoAI: Towards Building Reliable AI Software for Real-world Applications

Abstract

Deep learning is a powerful technique for many real- world problems. However, due to its unexplainable characteristic and over-fitting issue, there remains a great challenge for building reliable system with deep learning modules. In this paper, we present the idea of LegoAI that aims to build reliable AI software with pluggable modules of different functionalities, such as ensemble for fault tolerance and anomaly detection for result validation. In particular, we have applied the idea to develop a real-world AI software for handwritten digit recognition and achieved promising results.