Machine learning has been introduced more than 40 years ago in interactive systems through speech recognition or computer vision. Since that, machine learning gained in interest in the scientific community involved in human- machine interaction and raised in the abstraction scale. It moved from fundamental signal processing to language understanding and generation, emotion and mood recogni- tion and even dialogue management or robotics control. So far, existing machine learning techniques have often been considered as a solution to some problems raised by inter- active systems. Yet, interaction is also the source of new challenges for machine learning and offers new interesting practical but also theoretical problems to solve. In this paper, we address these challenges and describe why research in machine learning and interactive systems should converge in the future.