Project Initiation on Deep Learning in Context
MaineSAIL, the Maine Software Agents/AI Lab directed by Prof. Roy Turner, has long focused on the role context and contextual knowledge plays in intelligent agents’ behavior. Recently, a project has begun to apply insights gained over the years to deep learning (neural network) agents. This project builds on preliminary work done almost two decades ago in the lab, long before deep learning developed into the research, industry, and cultural juggernaut it is currently.
We are interested in deep learning systems that can learn in a variety of contexts, for example, one that can determine whether an object seen by an autonomous underwater vehicle’s (AUV’s) sonar is a mine or a harmless object: the same sonar information will mean something different near its base than in harbor in a hostile country. A problem with cross-context learning in deep learning systems is that the networks need to be large and the training consequently takes a long time and requires a great many training examples if the network is to avoid learning in one context being “washed out” by learning in another.
In this project, Turner explores the idea of learning and performing in-context. An agent would be equipped with a context manager (knowledge-based, neural-based, or a combination) that would be tasked with always maintaining a good assessment of the agent’s context, even as the situation changes. Deep learning about aspects of how to interpret the world and behave would be learned while in the context. When the context manager determines that the context has changed, the topology and weights of the network would be stored along with other information known about the context. Later, when encountering an instance of the context again, the deep learning networks would be re-instantiated.
Learning in context like this should help the agent behave more appropriately for its contexts with smaller networks, and it should reduce the training time and number of training examples needed.
The project is still in the early stages. Turner is actively seeking funding and students who might be interested in working on it.