Biologically-Inspired Hierarchical Learning Architectures for Image Classification

This thesis focuses on the topics of biologically inspired hierarchical machine learning methods for object classification. Mimicking the human brain to achieve human-level cognition performance has been a core challenge in artificial intelligence research for decades. Humans are very efficient in capturing the most important information while being exposed to a plethora of different stimuli, a capability that is used to represent and understand their surroundings in a concise fashion. Think about a kid that learns how to categorize objects through, either labeled or unlabeled, samples. It is a matter of fact that he/she is able to grasp the object concept by processing a few samples. This strong evidence highlights the existence of an extremely complex and engineered cortical mechanisms that allows such efficient learning. This Thesis draw inspiration from the fascinating biological brain organisms and its ability to learn simple concepts as well as complex notions from a few samples to introduce novel hierarchical learning models. It also grounds on the belief that the study of the vision sensory domain can provide a uniquely concrete grasp on the relevant theoretical and practical dimensions of the problem of learning in hierarchies. Thus, this Thesis provides an in-depth investigation of biologically-inspired hierarchical learning architectures for image classification. Pivoting on the belief that decomposability of the sensory data is a fundamental principle for learning good representations, the proposed hierarchical learning architectures adhere to such a property by aggregating simple features into more and more complex patterns as the structure becomes deeper and deeper. The underlying idea shared by these models is is to finally provide a different --artificial-- hierarchy of computations that mimics the human brain by abstracting away from existing highly-"engineered" models that are quite in vogue (e.g., Deep Neural Networks). To support each approach, experimental results on public datasets are conducted. Results demonstrate that exploitation of subsequent filtering and pooling strategies are the main ingredients of a hierarchical architecture able to build meaningful data representation.