Paper Title
Tiny Machine Learning Hardware Platform Computation Optimization

Abstract
Convolution neural network (CNN) is used in many artificial intelligence (AI) applications, including voice command recognition of mobile device. Tiny Machine Learning (ML) hardware is designed to reduce computational complexity. Tiny ML implementation with block floating point (BFP) matching error tolerance of neural computing is introduced in this paper to reduce the processing and memory storage of the system. The processing time can be reduced by 50% by adopting hardware accelerators. Furthermore, the recognition results obtained from applying BFP operations are promising, achieving over 0.97recognition accuracy, with a relatively small accuracy degradation of less than 1%. Keywords - Block floating point, Tiny Machine Learning, Convolution neural network.