Machine Learning and AI with Simple Python and Matlab Scripts

7Convolutional Neural Networks

7.1 Introduction

It is understood that when humans process visual information, they condense a broader visual field of view into multiple overlapping smaller receptive fields that are controlled by respective neurons in different areas of the brain (i.e. visual cortices). In order to recognize an object, a visual input is processed through hierarchically‐connected receptive fields. Neurons that control receptive fields respond to relatively small and simple visual stimuli, whereas neurons associated with receptive fields located deeper in the brain process increasingly larger and more complex optical features. As visual information progresses along the hierarchically connected receptive fields, neurons begin to represent visual objects in progressively more accurate ways until an object can be recognized by the brain (Figure 7.1a).

Figure 7.1 Feature recognition by the visual hierarchy of brain cortices (adapted from [70] / with permission of Grace Lindsay).

Figure 7.1b shows the growth of receptive field sizes as the brain processes visual stimuli by passing them from the primary visual cortex (called V1) through secondary visual cortices (called V2 and V4) and, finally, recognizing the object at its inferior temporal (IT) cortex. Figure 7.1c shows how neurons process an image as the brain processes it through the layers, each with increasingly ...