Deep Aggregation Autoencoders (Ma)
Topic for a bachelor/master's thesis
A fuzzy pattern tree (FPT) is a specific type of hierarchical fuzzy model . More specifically, the structure of an FPT is that of a binary tree, whose inner nodes are marked with generalized (fuzzy) logical and arithmetic operators, while the leaf nodes are associated with (unary) fuzzy predicates on a given set of input attributes.
Fuzzy pattern trees can be useful for a variety of machine learning applications. They can be used as regression or classification models with some interesting properties in terms of interpretability and the ability to incorporate expert knowledge into the modeling and learning process. Fuzzy pattern trees share a common property with modern neural network architectures : both model classes can be referred to as deep in the sense that raw values given by input attributes are successively combined (aggregated) with each other, building a feature-hierarchy layer by layer, eventually ending up with a final score in the output layer or root node, respectively. However, while fuzzy pattern trees so far have only be considered as binary trees, neural networks usually encompass a huge variety of more general (often fully connected) network structures.
Of special interest for many applications are the so-called autoencoder  architectures. These are frequently used to pre-train a compressed general purpose representation of the raw input features for later use in more specialized networks solving a concrete learning task. The idea of this thesis is to design a new deep aggregation autoencoder architecture that uses fuzzy logical operators and predicates like an FPT does while consisting of potentially fully connected layers compressing the input in a similar way as autoencoders. To accomplish this task, a new learning algorithm has to be designed.
Machine learning, programming, fuzzy logic (optimal).
-  R. Senge, E. Hüllermeier. Top-Down Induction of Fuzzy Pattern Trees. IEEE Transactions on Fuzzy Systems, 2011.
-  Ian Goodfellow and Yoshua Bengio and Aaron Courville. Deep Learning. MIT Press, 2016.