Hierarchical self-organizing learning systems for embodied intelligence

In this work, a framework of designing embodied intelligence (EI), along with the essential elements and their design principles, is proposed. This work intends to deploy the following design principles. Firstly, hierarchical self-organizing learning systems in the form of network made of neurons are the essential elements for building machine intelligence. The supervised, unsupervised and reinforcement learning are all necessary aspects of learning and are studied for machine intelligence building. In supervised learning, an efficient learning method for hierarchical multi-layered network structure is proposed and studied. In addition, a quantitative measure is proposed to quantify overfitting of a network in a given learning problem to determine proper network structure or proper learning period. In unsupervised learning, a sparsely-connected hierarchical network is developed to build the neural representations effectively and efficiently for densely-coded sensory inputs, and to enable the memory with large memory capacity and great fault tolerance. Secondly, the memory-based intelligence is not only for passive information processing and pattern storage. One of the critical capabilities of intelligence is continuous and intentional learning. Therefore, a goal creation system (GCS), also as a type of hierarchical self-organizing learning system based on simple and uniform structure, is presented that acts as the trigger for the agent's goal creation, memory management, active interaction and goal-oriented learning. As a self-organizing structure, it is responsible for evaluating actions according to goals, stimulating the learning of useful associations and representations for sensory inputs and motor outputs. It enables the more powerful hierarchical reinforcement learning, finds the ontology among sensory objects, creates the needs, and affects the agent's attention and perception. Biologically inspired structural design concept and the framework of EI proposed in this dissertation create a promising direction in the field of EI. It enables the desired capabilities for an intelligent machine to have, including the efficient, continuous and intentional learning, large representative memory capacity, and goal-oriented perception and action. The hierarchical self-organizing learning systems include all ingredients necessary to develop intelligence, and to motivate a machine to act on its own in its environment. Having the framework defined and design principles prepared, the future work will be done more consistently.