CCF: SHF: Small: Transformer synthesis

点击翻译按钮获取中文摘要

Just within four years of being first proposed, transformers have had a dramatic impact on the natural language processing (NLP) field and are also beginning to have an impact on other fields, such as computer vision. This success has largely been driven by large-scale pre-training datasets, increasing computational power, and robust training techniques. However, a major challenge that remains is efficient optimal transformer model synthesis for a specific task and set of user requirements. However, this is not easy to do since the design space of transformer models is vast. This project addresses this challenge through the development of transformer-synthesis methodologies and tools. Given the importance of transformers, such tools are likely to have a transformative impact on many application areas. The research will be disseminated to industry via tech transfer e.g., via open-source online distribution of source codes, summer internships, and by leveraging PIs involvement with local companies. Outreach and curriculum development plans will also be undertaken within the context of the proposed research.There is currently no universal framework that can navigate the vast transformer hyperparameter design space. Previously proposed transformer models are homogeneous in terms of data flow through the network. Unfortunately, this leads to very suboptimal transformer architectures. This project expands the transformer design space to incorporate heterogeneous architectures that venture beyond self-attention by employing other operations like convolutions and linear transforms. It will also explore novel projection layers and positional encodings to make hidden sizes flexible across various transformer layers. It will use a dense embedding to capture model similarity to significantly enhance search efficiency. It will develop a heteroscedastic surrogate model to further speed up search. It will include operations that optimize long-range interactions for long input sequences. It will also explore skipped connections and block-level grow-and-prune synthesis to improve architectural search efficiency.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.