Incremental Learning with Self-Organizing Bayesian Adaptive Incremental Network (SOBAIN)

Abstract

Neural networks and artificial intelligence have revolutionized how machines learn by mimicking aspects of human cognition. One key area in this field is the ability of agents to understand and imitate actions involving various objects, allowing them to pick up new skills by observing others. Understanding and imitating how others interact with different objects has become a big topic since it allows learning new skills by simply watching others. By constantly updating their own knowledge, lifelong learners can build on what they know over time. But the environments that artificial agents deal with are very different. Existing models designed for “lifelong learning” typically work with simplified experiments and datasets made up of static images, which limits their effectiveness for real-world applications. In this study, we propose a developmental model focused on how agents can learn about objects and actions through sensorimotor feedback, enabling humanoid robots to mimic actions more naturally. Our approach, "SOBAIN," works in three stages: neuron activation, neuron matching, and neuron learning. First, neurons activate based on specific traits that determine if they should "fire" or not. Then, we match the best neuron by comparing activation levels. At each learning point, new neurons connect to the network to match learned data. The final stage uses this network to refine and apply learned information, helping address memory loss issues in lifelong learning. By using these techniques, SOBAIN helps robots adapt their upper body movements in line with the sensor's feedback, creating a chain of neurons that builds over time as the robot learns new actions.