Brain modeling is a research area within computer science devoted to the study of complex and dynamic computing algorithms that imitate brain function regarding the information processing properties of the structures that make up the nervous system. The computational and mathematical structures are composed of interacting modules, whose coordination aims to enhance their problem-solving capabilities. The computational models of the visual cortex use non-trivial interactions between a large number of components. In this paper, we propose a hierarchical structure that mimics the information flow and transformations that take place in the human brain. This paper describes a virtual system composed of an artificial dorsal pathway-or ''where'' stream-and an artificial ventral pathway-or ''what'' stream-both are fused to recreate an artificial visual cortex. In previous work, the model was refined through genetic programming to enhance its performance over challenging object recognition tasks. The system finds good solutions during the initial stage of the genetic and evolutionary search. In this paper, the goal is to show that a random search can discover numerous heterogeneous functions that are applied to a hierarchical structure of our virtual brain. Thus, the proposal presents two key ideas: 1) the concept of function composition in combination with a hierarchical structure leads to outstanding object recognition programs, and; 2) multiple random runs of the search process can discover optimal functions. The experimental results provide evidence that high recognition rates could be achieved in well-known object categorization problems; consequently, this paper corroborates the importance of the hierarchical computational structure described in the neuroscience literature. INDEX TERMS Automatic programming, brain modeling, artificial visual cortex, brain-inspired computing, heuristic computing, deep genetic programming.

No abstract

This work describes the use of brain programming applied to the categorization problem of art media. The art categorization problem—from the standpoint of materials and techniques used by artists—presents a challenging task and is considered an open research area. Brain programming is a machine learning methodology successfully tested for the problem of object categorization; however, when working with art images, the objects in pictures of the same category may be different from each other regarding image content. Therefore, it is necessary to find the best set of functions that extract specific features to identify patterns among different techniques. In this study, we show a comparison with deep learning to understand the limits and benefits of our approach. We train and validate solutions with the Kaggle database and test the best results with the WikiArt database. The results confirm that brain programming matches or surpasses deep learning in three out of five classes (over 90%) while being close (less than 5%) in the remaining two with significantly simpler programs.

No abstract

No abstract