|Speaker :||Shirin Jalali|
|Nokia Bell Labs (Murray Hill)|
|Time:||3:00 pm - 4:00 pm|
|Location:||Paris-Rennes Room (EIT Digital)|
Solving most inference tasks, such as denoising and linear regression, relies on exploiting the structure of the desired class of signals (e.g. images). Traditionally, such structures are discovered by domain experts after extensive studies. This has to a great extent limited the discovery and application of complex structures that exist in many signals of interest. Learning-based methods that automatically recover the structure of the source from available training datasets provide a promising alternative solution. However, for continuous-valued signals, learning the source distribution is extremely challenging, and theoretically-founded computationally-feasible approaches are yet to be found. In this talk, I will discuss recently-proposed structure learning methods that, instead of learning the full distribution of the source, learn its key features that are relevant to solving inference problems. As I will discuss, this method, while substantially reducing the computational complexity of the structure learning task, leads to asymptotically optimal learning-based estimators.