This artificial intelligence (AI) system is designed to understand comparative relationships between input classes, allowing it to predict and reason about the nature of an input belonging to a unknown category by relating it with the memorized past relationships. The ability of AI to learn has improved dramatically in recent years, thanks to a greater understanding of human neural networks, but it is still nowhere near the ability of a human. Existing AI designs require a large number of training samples to build a classification system and learn proper outputs for future, undetermined inputs. Deep learning, a method which uses layered neural networks, has improved such AI design. However, it lacks the human ability to flexibly categorize and respond to inputs from unknown classes. Researchers at the University of Florida have developed an AI system that unifies deep learning and reasoning. It not only predicts classification outputs from previously known categories, but also characterizes inputs from unknown classes by relating them to an explicit structured memory bank of past instances.
AI system that can produce proper outputs for previously unseen inputs by learning through comparative relationships
This AI program is a spatially forked model containing two primary components: primary learner and a comparator network. The primary learning network is a conventional deep learning system, while the comparator uses a Siamese style network design. The primary learner has an adjacent structured memory bank that predicts the outputs from a given input and also relates inputs to all past memorized instances, assisting in creative understanding. The comparator network aids in determining the relationship between classes. This type of analysis unifies deep learning and reasoning, allowing the AI design to use classes to categorize examples, just as a human does.