Imagine an automated assistant that actually understands your goals and job, and can be customized to the way you want things to be done. If we understand the computational processes underlying training Artificial Intelligence (AI) systems, we may well learn how to educate and train humans, and develop AI systems that can then educate us. Creating AI systems that understand our needs and learn new tasks may further help us develop systems that allow the elderly to maintain more autonomy and better and even cheaper home care. If we have intelligent home robots, we can eliminate the drudgery of home maintenance - much more than just automated vacuums. Therefore, Dr. John Laird and his team at the University of Michigan are trying to understand how the mind works from a computational perspective, having embarked on a long-term commitment to creating a highly functional, integrated cognitive architecture named Soar that has many, if not most of the capabilities of humans, including perception, decision making, reasoning, problem solving, language, and many forms of learning. The more we can get AI systems to help us make intelligent decisions, the better prepared we will be for dealing with our everyday problems as well as those of our broader society.

Inspired mainly by psychology and marginally by neuroscience, Dr. Laird is driven by the desire to understand the computational structure of the mind - or its cognitive architecture - and using that understanding to build general intelligent systems. The hypothesis is that there is a level of description of the mind that can be understood in terms of structures such as decision making, short-term memory, and different types of long-term memory. Dr. Laird and his team aim to understand and build systems that have semantic memory (memories of facts), episodic memory (memory of experiences), and procedural memory (skills - how to do things). Not only is it about the different components - such as how memories are stored and retrieved, but it is also how all of the pieces work together to produce coherent, adaptive behavior. This theory of intelligence and the mind is embodied in Soar, a computational cognitive architecture, which has been in development for over 30 years and is used worldwide for developing AI agents and applications. By keeping his work open source and freely available on the web to the broadest community possible, Dr. Laird hopes to enable programs that will widely benefit the society.

Dr. Laird’s research has two main areas of focus:

• Extending Soar’s Capabilities: By looking for the kind of novel capabilities that humans have that AI systems don’t have, Dr. Laird and his team can understand what is necessary to achieve those capabilities and extend the capabilities with Soar. For example, over the last ten years they have added mental imagery, semantic memory, reinforcement learning, episodic memory, and limited aspects of affective processing (emotion) to Soar. Dr. Laird and team are currently refining Soar's long-term memory systems inspired by psychological research on prospective memory, creating memories to be used in the future, such as remembering to pick up some milk on the way home.

• Interactive Task Learning: Dr. Laird and his team are also working to use the existing architecture to explore a complex cognitive capability that draws on many of the architectural mechanisms. Their current focus is on Interactive Task Learning - enabling the software to learn a new task through natural interaction with a teacher. What makes this exciting is that it has the potential to eliminate programming; if we can teach our AI systems new tasks by talking to them and instructing them, we will no longer have to rely on AI researchers to pre-program them with all of their capabilities. Dr. Laird and his team have already demonstrated the ability to teach the system hierarchical tasks, simple games, and puzzles through restricted natural language instruction. In this case, Soar is embodied in a tabletop robot that uses a Kinect vision system and a robot arm to manipulate blocks that are used in its tasks. Interactive Task Learning is very challenging because there is no way to pre-specify the knowledge for how to perform a task and instead, its agent must learn that through natural language instruction from a human. However, if we can achieve it, it promises to greatly simplify the way we interact and teach computer systems.