DeepMind Looking to Implement AI with the Way Children Learn

Any parent will tell you that most toddlers will try and get their hands on anything within reach, and more often than not, will try and taste whatever they have in their hands. While those actions may seem goofy, it's a method kids use to explore the things around them, which in itself is a form of learning. It's thought that this form of early learning supports a more robust generalization and intelligent behavior later in life.

Researchers at DeepMind want to incorporate that early form of learning to AI agents, where it could then cte new exploration techniques for powerful, abstract task generalization, which they could benefit from. For example, when infants go mobile and begin to crawl, the exploration of their surrounding area appears to show them learning about objects and spaces. Even preschoolers who play with a toy develop a theory about how the toy functions, like associating how blocks work based on their color, and then taking that experience and applying it to a new toy they've never seen.

AI agents are already capable of performing the same feat, but struggles without some form of human oversight or intervention. The researcher's approach to bridging the gap between child exploration and machine learning utilizes the DeepMind Lab — a Quake-based learning environment for agent-based AI learning involving first-person navigation and puzzle-solving tasks. Those tasks involve physical or spatial navigation skills and are modeled after children's games.

In a unique setup, kids interact with the DeepMind Lab via a custom Arduino controller, which offers the same four actions AI agents would use in that environment, including moving forward, backward, left, and right. Experiments were designed by the researchers to determine two things- whether differences in child exploration exist within unknown environments, and whether kids are less susceptible to responding to goal-oriented exploration over AI agents.

In one test, kids were told they could explore a maze all on their own, while a second had them search for a "gummy." The researchers found that the kids who could freely move within the maze, the children's strategies closely resembled those of an AI agent 89.61% of the time, while the goal-orientation gummy test showed kids made choices consistent with AI agents 96.04% of the time.

Another test broken into three phases, with the first exploring a maze with no conditions, the second to explore areas in sparse conditions with a goal but no reward, and the third providing a dense condition with both a goal and reward. The collected data suggests children are less likely to explore an area in the dense/reward condition, which doesn't seem to have an impact on the children's performance in the final phase. The same isn't true for AI agents, as dense/rewards make them less incentivized to explore, which leads to weak generalizations.

The collected data from the experiments will lead to new questions on AI agent learning, and according to the researchers, "In asking these questions, we will be able to acquire a deeper understanding of the way that children and agents explore novel environments, and how to close the gap between them."