Description:
In this talk I will describe two design problems in areas of chemistry and computer science which yield themselves to machine learning techniques.
The area of supra-molecular chemistry deals with mechanisms of noncovalent assembly of particles. The interest in understanding the underlying mechanisms is two-fold: it provides insight into protein complex formation and paves the way for application of these mechanisms in nanotechnology. The spontaneous assembly of particles is referred to as self-assembly and the ability to design such processes holds great promise for nanotechnology. The problem of design of self-assembly processes turns out to be a task surprisingly familiar to the machine learning community: that of probability maximization. The standard Boltzmann machine learning rule can be applied to this task, and I will demonstrate a way to speed up the evaluation of the derivatives via importance sampling. In addition, I will demonstrate several methods for evaluating the probability of a shape under a self-assembly process.
The second part of my talk will be devoted to another problem which at first blush does not admit a probabilistic interpretation. This is the problem of inferring programs given input/output pairs. I will introduce a probabilistic representation of code. This representation gives rise to a distribution of state sequences and allows approximate inference in form of loopy belief propagation. I will illustrate the performance of this method on tasks of discovering programs for polynomial computation and list reversal given only examples of the input/output pairs.
Speaker(s):
Vladimir Jojic, Ph.D. candidate, Computer Science, University of Toronto; Microsoft fellow
|
