USING TENSORFLOW ECOSYSTEM OF TOOLS FOR IMPLEMENTING NEURAL NETWORKS ON CYPRESS SEMICONDUCTOR MICROCONTROLERS

Abstract

Today, a lot of difficult tasks require the implementation of neural network methods. According to the definition, neural networks are used to recognize the correlation between input and output data, that cannot be easily formalized by a programmer. For example, the programmer cannot formalize basic relationships between image pixels and recognizable items on the picture. In this case, the traditional programming paradigm is replaced by machine learning.

The problem of using neural networks on difficult tasks is a large amount of data. Also, the process of using neural networks requires a lot of floating-point operations. Usually, a neural network task requires a real-time response, and for this reason, neural networks use cloud services. Cloud services get input data from an IoT peripheral, process it, and send the response of output data back to an IoT peripheral. Such a way of neural network usage organization with IoT peripherals has some disadvantages. For instance, an IoT peripheral should have a connection to a network, personal user data can be stolen due to transporting, cloud service should be powerful and has a multithreading module for real-time response to a large number of the devices at the same time.

An alternative method of using neural networks on IoT peripherals is quantization. Quantization is a technique that reduces the size of neural network weight by converting it to a smaller data type. Using integer quantized values boosts the speed of calculation and reduces the size of the whole neural network. This method avoids a connection to a powerful cloud service, so personal data do not leave an IoT peripheral and cannot be stolen.

Key words: neural networks, quantization, machine learning, neural networks on IoT peripherals and microcontrollers.