Neural Networks for Higher-Order Reinforcement Learning and Multi-Perspective Generative Modeling

Abstract

What does it mean for a system or machine to exhibit intelligence? In machine learning, a system is said to exhibit artificial intelligence provided that it improves its performance with experience on a given task. However, some tasks require more cognitive load than others, and higher-order cognitive processes are needed to solve some tasks effectively. We developed a framework inspired by how the human brain utilizes working memory to solve complex tasks, and constructed a model with the ability to solve tasks with multiple context layers. Our results show that working memory plays a vital role in solving cognitive context processing tasks. Furthermore, we extended our model to utilize the generalization benefits of output gating within a working memory system. This modification resulted in successfully transferring knowledge across a temporally extended, partially observable grid-world maze task that required the agent to learn three tasks throughout the training period. Finally, Generative Adversarial Networks (GANs) can be viewed as using higher-order cognitive processes to perform image-to-image translation. More specifically, we are concerned with generating video frames from a missing camera feed, where each feed is viewing the same area from different positions and angles. Results show that our model can produce realistic video frames that resemble the missing video source.