Contextual learning for interruption tolerance and multitasking

Abstract

Natural environments abound in event streams that require interruption tolerance and multitasking. This project introduces Mandala, a neural network that improves multitasking in the form of dealing with intervening events from overlaid causation chains, a capability that a conventional recurrent artificial neural network (RNN) struggles with, as evidenced by the results. This capability is also tolerant of interrupting events. Mandala achieves this by accumulating contextual tiers of temporal states that are fed into a multilayer perceptron (MLP) at each time step. Mandala is also an effort to combine the Morphognosis and Mona neural network models into a comprehensive model for learning and behavior. Mona features a contextual causation learning with goal-directed motivation. Morphognosis features contextual MLP learning. In addition, accumulating temporal information discretely labels hierarchical cause-and-effect relationships that can be used for augmented processing. In the case of Mona, channeling motivation through the network for the purpose of goal-seeking requires this feature.

Keywords