I specialize in understanding the connections between components of a system.
I did my thesis studying compression of Gibbs fields with David Neuhoff at the University of Michigan. Data compression is excellent training for developing systems, in particular machine learning systems. For example, three essential components of a machine learning system are a component for learning a model; a component for computing probabilities with respect to the model; and a component for taking action based on a computed probability. These are also the three essential components of a data compression system, for example the artihmetic encoding based systems that Prof. Neuhoff and I have been developing for Gibbs fields. In such a system, the action component is the assignment of a bit string to a piece of input data, based upon the probability computed for that piece of input data, using a previously learned model.
Moreover, my thesis focused on data compression for sources of data modeled by Gibbs distributions, which are not only a very general class of models for multivariate random variables, but one whose parametrization in terms of inter-connectedness between random variables makes them especially appealing for data corresponding to social and device networks. Indeed, I have recently introduced A Marketing Game as a model for the competition between companies to use marketing to influence the socially-contingent decision-making of consumers. Two essential components of this model are the use of social network data to learn network influences that determine consumer choice, and predictions of market share resulting from candidate marketing allocations applied to a network with such influences.
In addition to experience with the design of systems involving learning, prediction, and action, a background in data compression provides another substantial benefit with regard to systems design. Data compression, or source coding, is one of the two initial fundamental branches of information theory, the rigorous study of the quantification of data storage and transmission. As such, many of the metrics used in machine learning have operational significance in data compression, giving me a strong understanding of how individual components of a system affect overall system performance.
Tell me about the problems you are trying to solve and what's at stake.
I will crunch some numbers, develop an initial plan.
We will discuss my ideas and move from there.
We will agree on a plan of attack, deliverables, and a timetable.
I will keep you apprised of my findings with periodic progress reports.
I will deliver a well-documented solution that both solves the problem and adds to your company's brain trust.
Ann Arbor, Michigan, United States