Feature selection and hyperparameters tuning of neural networks applied for wood quality classification

• Grant number: 18/02822-1
• Support Opportunities: Scholarships in Brazil - Scientific Initiation
• Start date: April 01, 2018
• End date: November 30, 2018
• Field of knowledge: Physical Sciences and Mathematics - Computer Science - Computing Methodologies and Techniques
• Principal Investigator: André Luis Debiaso Rossi
• Grantee: Mateus Roder
• Host Institution: Universidade Estadual Paulista (UNESP). Campus Experimental de Itapeva. Itapeva , SP, Brazil

Abstract

Machine Learning Techniques (ML) have been widely used in the most diverse areas of knowledge for the induction of descriptive and predictive models. Many of these techniques have achieved great success when applied to specific tasks, such as the problem of sorting wood quality through images. This problem has been studied by the proponent of this project. Despite the satisfactory results, the research continues to advance in order to improve the predictive performance of the proposed system. For this, one of the tasks investigated is the adjustment of the hyperparameters of AM techniques. These hyperparameters directly influence the induction of the models and must be adjusted for each set of data. Meta-heuristics based on evolutionary process and swarm intelligence, such as Particle Swarm Optimization (PSO), have been employed for this purpose. These have specific mechanisms to escape the local minimums, which are very common in the hyperparameter adjustment problem. In this project we will investigate the use of the PSO meta-heuristic to adjust the values of the hyperparameters of Artificial Neural Networks (ANNs) concomitant with the selection of attributes for the problem of wood quality classification. The selection of attributes may be necessary, since different methods will be used to extract relevant characteristics from the wood images. The characterization of the images also has great influence on the predictive performance of the ML techniques. These techniques deal with the transformation of images into numerical values that are useful for use with AM techniques. Some methods of extracting the most commonly used features will be investigated, such as measurements obtained from the co-occurrence matrix of the gray scale and the local binary pattern. It is expected that the PSO will be able to find the best set of hyperparameters of the RNAs and the best set of attributes describing the images will be selected, increasing the accuracy rate of the ML technique. (AU)