Beyond supervised pattern recognition: Efficient learning with latent combinatorial structure
June 15, 2016, 11AM
Supervised pattern recognition with 10^6 training data and 10^9 layered parameters has brought tremendous advances in artificial intelligence. However, there are two main limitations to this approach: 1) The knowledge learned in one area doesn't easily transfer to another areas and 2) supervising every single task is not only infeasible but also requires huge amounts of human labeled data which is costly and time consuming. In this talk, I will suggest a unifying framework which jointly reasons the prediction variable and the underlying latent combinatorial structure of the problem as a way to address such limitations. To demonstrate the practical benefits of the approach, we explore classification, localization, clustering, and retrieval tasks under settings that go beyond fully supervised pattern recognition..