Training and Analysis of Hyperparameters in Neural Networks for Computer Vision Applications: A Didactic Approach

Machine Learning models are known for having millions of parameters considering different applications, such as object recognition and self-driving cars. To avoid spending more time than necessary in the training stage and more importantly, to achieve a satisfactory model, the hyperparameters should be correctly defined. The main goal of the proposed project was to identify and evaluate the influence of hyperparameters in the training of neural networks for computer vision applications. Besides that, we aim to provide a didactic approach to analyze the acquired generalization capability of a model and easy visualization of the impact of different parameters during the training phase.

I. Introduction

It is well known that Machine Learning is one of the main fields of research in the current days, with the number of projects and research papers increasing each year, as shown in Fig. 1, which represents the number of papers submitted in a timeframe of eight years in one of the most important conferences in the field.