Emotions on Edge - the Dependence of Different Characteristics of the Convolutional Neural Network on the Number of Classes

Machine learning is most often associated with powerful computers, but lately, it is increasingly present on both microcontrollers and systems on a chip. Initially, simple machine learning algorithms were used on these platforms, but today we already see high-resolution cameras and convolutional neural networks used to process camera data. Models for such systems are prepared on large and powerful computers, but previously trained neural networks could be copied and used on microcontrollers and systems on a chip. This paper aims to measure the performance and characteristics of systems on a chip in which a convolution neural network for recognizing facial emotions is implemented. A different number of emotions are used to determine the impact of this change on performance, and the effect of quantization on performance and model size is also analyzed.