My full-time role is in the development of Infer.NET which is a .NET platform for machine learning. I am also involved in its support within the internal and external communities, and its application to various problem domains. Infer.NET uses a model-based approach to provide state-of-the-art algorithms for probabilistic inference from data. Various Bayesian models such as Bayes Point Machine classifiers, TrueSkill matchmaking, hidden Markov models, Bayesian networks, as well as a large range of custom models can be implemented using Infer.NET. You can think of Infer.NET as a language for modeling uncertainty (though in practice it is an API). Rather than requiring data to be shoe-horned into existing black-box machine learning algorithms, it encourages explicit modeling of the processes that generated the data. It then provides a compiler to convert inference queries on the model to a tailored bit of algorithmic code.
Before working at Microsoft Research, I spent several years developing Advanced Control software at Aspen Technology. One of the highlights was the development of a fully non-linear Model Predictive Controller (Aspen Apollo) which has been widely adopted in the polymer manufacturing industry. This was made possible by developing a new form of non-linear regression model (a Bounded Derivative Network) which has more natural interpolation and extrapolation properties than a traditional Multi-layer Perceptron model, and which provides guaranteed global behaviours such as montonic responses in specified input variables. A description of Bounded Derivative Networks can be found in ‘Introducing the bounded derivative network—superseding the application of neural networks in control’, P. Turner, J. Guiver, Journal of Process Control, pages 407–415, 2005. A description of the many practical issues faced in putting a non-linear controller into a large-scale manufacturing plant can be found in ‘Experiences with Non-linear MPC in Polymer Manufacturing’, Kelvin Naidoo, John Guiver, Paul Turner, Mike Keenan, Michael Harmse in Assessment and Future Directions of Nonlinear Model Predictive Control (Lecture Notes in Control and Information Sciences)
The Aspen Apollo controller was derived from an earlier controller developed at NeuralWare (‘A nonlinear industrial model predictive controller using integrated PLS and neural net state-space model’, H. Zhao, J. Guiver, R. Neelakantan, L.T. Biegler, Control Engineering Practice, 9, 2001., ‘State space nonlinear process modeling: identification and universality’ ,G.B. Sentoni, J.P. Guiver, H. Zhao, and L.T. Biegler. AIChE Journal, March 1998). At NeuralWare (one of the early commercial neural net companies, founded by Casey and Jane Klimasauskas in 1987), I was a principal research engineer and was, for many years, the lead sofware developer, responsible for the design and evolution of the company’s neural net products.