Learning Systems Group

The Learning Systems group works on the theory, algorithms, and applications of automated learning. Our goal is to understand the principles of automated learning and to develop state-of-the-art learning solutions for real-life problems.

Learning can be viewed as an alternative approach to system design. Instead of the conventional way of mathematically modeling the task at hand and implementing the model as a system (a computer program or a piece of hardware), we let the learning process do the work for us. We start with a generic model, such as a neural network, that has a number of "untuned" internal parameters. Depending on how the parameters are tuned, the model can implement vastly different tasks. The role of learning is to take examples of the task, such as inputs together with their target outputs, and use this information to tune the parameters of the learning model to mimic the desired task.

The learning approach is essential when the task at hand does not lend itself to exact mathematical modeling, which is the case in many practical applications. Successful stories of the learning systems can be found in vastly different domains, from medical to financial to industrial.

Some of the highlights of our group's work have been the use of hints in learning, the theory of learning from very noisy data with application to Computational Finance, and the application of learning to medical image recognition. We have on-going work on a number of theoretical and algorithmic fronts, including data pruning, monotonic learning, infinite model aggregation, and the bin model for generalization.