Совместный семинар ИППИ РАН и Сколтеха
11 октября (вторник), 1830, аудитория 615 ИППИ РАН
Ермек Капушев (Сколтех, ИППИ)
Deep Gaussian Processes
Gaussian processes is a powerful tool for regression and classification. However, for more complex and abstract data sets deep architectures have proven to be more efficient. In the seminar we will discuss how GP-based deep architecture can be implemented. We will consider different approaches to learning model parameters: expectation propagation, variational inference and its autoencoded version. The resulting model is no longer a Gaussian Process, but allows to learn much more complex interactions between data. Such fully Bayesian treatment allows to model complicated data even when the data set is scarce, in contrast to deep networks which usually require large data sets. Some applications of the model to artificial and real data sets will also be considered.
11.10.2016 | |