In our view, AI simply means that computing systems perform complex tasks on real-world signals effectively, and move beyond what traditionally programmed computers have done.

Machine learning has so far proven itself to be the best way to achieve artificial intelligence like capabilities. Since humans are unable to directly program such complex tasks, we can instead employ mathematical optimization to produce programs in the form of learnt models.

Much of machine learning involves continuous optimization (e.g., stochastic gradient descent), but many problems are inherently discrete, such as feature and data subset selection, structure learning, summarization, coresets, sketching, sparse models and so on. Submodularity and supermodularity is a natural and fast paradigm to study hard discrete optimization problems and apply them to real-world tasks.

Speech and language are amazing natural signals that humans easily produce and consume. How can we get computers to do the same?

Graphical models are a formal, mathematical, but visually concise and intuitive way to visualize important properties of families of probability distributions. They also facilitate the development of practical and scalable strategies to compute probabilistic inference overs such families for many machine learning problems.

Scientists at the Melodi lab work on other problems in computational biology (genomics, proteomics), computer vision, sensor networks, network computing, smart cities, the science of data management, computational music, high-performance computing, and so on. See our publications list for more.