What Can Babies Teach Us About Making Robots?
February 15, 2018.
How long before robots take their mastery of games like chess or go and generalize to other arenas of engagement on Earth? Some scientists, namely developmental roboticists, explore the underlying mechanisms which could allow lifelong, open-ended learning, similar to learning found in humans. Robot’s ability to learn from their own experience holds the key to their eventual realization of intelligence which surpasses humans.
One emerging theory about learning in humans posits that active learning involves comparing our expectations about the world to the actual world as we experience it. Learning itself can be considered a simple update of our expectations about the world to match the reality of experience as it unfolds. Expectations about the world are a vital component to how humans manage to survive and learn. For robots to have the correct expectations and be able to generate new expectations on-the-fly, they have to be able to learn to build expectations from scratch. Is there a better model system of an active learner than the human infant?
One initial attempt at creating baby-like robots, in order to explore how we might program open-ended learning into machines, was conducted by Dr. Pierre-Yves Oudeyer, a former computer research scientist for Sony, now the director of the Ensta-ParisTech FLOWERS team in France. This study attempted to test a model of learning that functions, in a way, by monitoring its own learning progress. If a task was too hard, the robots would not pursue the task. If the task was too easy, the robot recognized the activity’s futility. If, however, the task was intermediately difficult and the robot could monitor its own learning progress, the robot would continue engaging with the activity. The robot could monitor its own ability to build better expectations about the outcomes of its own engagement with a task. Perhaps, by utilizing the framework outlined by Oudeyer for self-organization in learning (papers found here, and here), we could begin to scale the open-endedness of learning in embodied machines up to human levels and beyond.
Steven Elmlinger, Psychology
Steven is interested in how perceptual intelligence (or lack thereof) gives rise to learning about an agent’s environment. His main focus is on the role of parent-child interaction in the communicative development of human infants.