| List |
| Topic: | Exploring Deep Declarative Neural Networks |
|---|---|
| Department: | Katedra kybernetiky |
| Supervisor: | Torsten Sattler |
| Announce as: | Diplomová práce, Bakalářská práce, Semestrální projekt |
| Description: | In standard feed-forward neural networks, each node computes a function of its input, where the parameters of this function are learnable. As
such, the function computed by these nodes is explicitly defined. Gould et al. recently proposed a novel set of nodes, named declarative nodes. Rather than explicitly defining a function, the output of a declarative node is a set of parameters that minimizes an optimization problem. This results in a new and powerful set of capabilities as it allows much more complex computations to happen in nodes. Gould et al. provide two examples of the use of declarative nodes: robust pooling unaffected by outliers (compared to simple mean pooling) and the projection onto hyper-spheres. Since deep declarative networks are a very recent concept, the goal of this thesis is to explore their capabilities on a set of illustrative examples. |
| Bibliography: | [Gould et al., Deep Declarative Networks: A New Hope, arXiv:1909:04866] |