Open-world link prediction via type-constraint embedding and hybrid attention for knowledge reuse of AI chip design

AI chips play a vital role in real-time AI applications. AI chip design requires vast knowledge reuse in integrated circuits and neural networks, and this knowledge continues to grow at a rapid rate, leading to a heavy burden on designers. Automated methods are urgently needed to handle the massive and rapidly growing literature. In the realm of traditional EDA, deep learning can produce excellent designs by learning from a large number of design examples. However, in front-end design such as algorithm and architecture design, a considerable amount of knowledge exists in the fast-growing literature. The use of this knowledge is still in a primitive state of manual processing. Therefore, this paper leverages the knowledge graph and the link prediction to organize and reuse AI chip design knowledge (ACDK). First, a knowledge graph is built to organize ACDK in triple format. Second, an open-world link prediction model named TC-THA is proposed, which aims at the closed-world limitation, the lack of type constraint, and the deep inner-association learning weakness of the related methods. TC-THA converts the type constraint of ACDK triples into vectors for better training and inference. It also uses a hybrid attention mechanism for the sufficient semantic interaction among inner-triple information, which is used to capture the deep internal associations of knowledge for the superior ACDK reuse. Experiments show that the TC-THA outperforms other methods and achieves the best MRR and score of 0.34 and 0.51 in ACDK link prediction. A case study of ACDK reuse shows that, for the given problems, the TC-THA predicts highly feasible solutions, which can be confirmed in the latest AI chip studies.