A synthesis flow for fast convolution unit based on molecular reactions

Nowadays, molecular computing and artificial intelligence have drawn intensive attentions from both academia and industry. Therefore, people begin to seek possible combination of these two promising areas. In this paper, with the aid of chemical reaction networks (CRNs), a design methodology of realizing the key block of fast convolution unit (FCU) in convolutional neural networks (CNNs) is proposed. In order to address the difficulty when “compiling” a complex traditional circuit into CRNs, this paper devotes itself in offering a systematic method to construct the n-parallel FCUs, where a 3-parallel FCU case is employed as a running example. Theoretical analysis, numerical simulation, and complexity comparison have proved the validity of this method and ensured its convenience, robustness, and efficiency. Though preliminary, the proposed method can be readily employed to design other arithmetic blocks of CNNs. Thanks to the high parallelism of CRNs, it is believed that this method offers another perspective to address the complexity of CNN implementation, especially when the system scale is large.