Socially Guided Machine Learning

Socially Guided Machine Learning (SG-ML) seeks to augment traditional computational learning systems by enabling them to interact with humans. The goal of SG-ML is to build more intelligent computational systems that can leverage the human capacity for social learning and, in turn, advance beyond the state-of-the-art in a variety of learning scenarios such as direct tutelage, learning from demonstration, social referencing, imitation learning, and more.

Traditional notions of Machine Learning often assume batch data, limited channels of feedback, and fixed notions of what role feedback can play. In contrast, humans are predisposed to engage in learning tasks using a variety of feedback strategies, across many modalities. If we want computational systems to be able to engage with humans in collaborative learning tasks, we must adapt machine algorithms to humans’ teaching styles.

Under SG-ML, learning is a fundamentally collaborative process in which the teacher scaffolds and guides the exploration of the learner, while the learner conveys their progress on understanding and learning the task. Both actively shape and tune the behaviors of the other to make the learning more efficient and effective. Although SG-ML systems are designed to take guidance from the teacher to accelerate learning when he/she is present, they also have internal motives for autonomous curiosity- and task-based exploration when the teacher is not present.

Socially Guided Machine Learning is a broad research agenda that encompasses many projects to realize this vision of computational agents that leverage human expectations and work collaboratively in order to learn more quickly, flexibly, and completely.

Publications

Socially Guided Machine Learning: Unlocking the Power of Human Teaching for Machine Learning