S9 SIIA Interactive Machine Learning


It merges machine learning and human-computer interaction. While traditional machine learning systems process the data that have been given to them in advance, this course considers that the learning process could benefit from interactions with the environment as well as with a human, and that inputs and outputs from and for humans carry meaningful information. Indeed humans may provide input to a learning algorithm, including inputs in the form of labels, demonstrations, advice, rewards or rankings. The interaction is all the more useful as the human can guide along the learning process while adapting his guidance to the outputs of the algorithm. This interaction can be in the form of feedforward or feedback information. The timing of these interactions can be preset, left to the teacher’s initiative or even to the learner’s initiative. In the latter case, the algorithm called “active learner" can decide when,about what, how and with whom to interact to optimise its learning process. Thus a bidirectional dialogue can emerge. Another point addressed by this course is the self-representation of the machine through its interaction with its environment. According to cognitive science, our skills and our representations of ‘a’ world come from our sensorimotor capabilities. These latter are the basis to a more abstract representation and some researches aims to propose models and algorithms that reproduce the underlying mechanisms. Indeed, these mechanisms could provide a mean to resolve the problem of the knowledge representations of unknown environment. The models allows for an anticipation of the perception from the action of the agent and for a construction of its own world. These researches are generally inspired by cognitive science.