Algorithm for Multi-Source Deep Transfer Learning

The basic architecture of a deep transfer algorithm with separation of concerns. The algorithm is feed some data D, has some bank of prior knowledge in D, and relies on two components: a standard machine learning algorithm to analyze the data, and an agent to build an informative prior and/or set the hyperparameters. In this particular case, supervised learning is done with Gaussian processes, but it could also use Support Vector Machines (setting the parameters) or any other supervised learning algorithm. We are looking for the best model given our data and our bank of prior data-sets. In this case, the agent's role is to establish the prior, i.e.: to create a bias toward more likely functions, and choose the hyper-parameters for Gaussian inference. Modelling with Gaussian processes requires a few free parameters (the hyperparameters) and the agent learn to select them. To search efficiently, the agent (reinforcement learning) will use natural language processing, read the labels in the data-sets, i.e.: x1 = humidity, x2 = Linux distribution, and learn to exploit this information to establish the best informative prior. Unlike other deep transfer algorithms like TAMAR, this approach can deal with an arbitrarily high number of sources and has no fixed method of performing transfer: it learns to do it. Reinforcement learning relies on rewards, in this case the reward will be established by the errors of the model during cross-validation and generalization, and how well it performs against a non-informative prior (if available). It should be possible to also tests agents against each other (i.e.: each with a different supervised learning algorithm). An important tool used to exploit the information in the label will be semantic clustering (unsupervised learning), which should clusters similar variables together and help the agent learn how to perform effective transfer.