P2: Network Principles for Building Deep Convolutional Computing Architectures

Project Description:


Deep layered convolutional neural networks form a basis for building Cognitive Computing architectures which can be trained and adapted to different sensory recognition tasks. In order to define and implement efficient scalable network architectures, two aspects need to be particularly investigated, namely appropriate design principles for the architectural layout and learning mechanisms for layers of hierarchical feature detectors and their combination. Regarding such principles, mechanisms of, e.g., reinforcement learning and evolutionary computation will be investigated for non-local optimization of network architectures and their components in conjunction with local learning of feature representations.

In this dissertation project the activation dynamics during information processing in recurrent convolutional networks will be investigated with a special focus on how information is integrated over time. To this end, canonical network principles will be identified and analyzed. For example, convolutional networks will utilize a simplified model of a cortical column (to serve as a computational unit or building block) and its interaction with other units in layered architectures of feedforward and feedback interaction. The individual representations of such layered networks will be learned by principles of hierarchical learning to build deep networks relying on principles of modified local Hebbian learning and global modulating reinforcer signals. An evaluation of such network architectures will utilize benchmark datasets, such as, e.g., MNIST, CIFAR, or object recognition challenges.

Direct supervisor:

Direct supervisor:

Expert advisors:

Affiliation:Ulm University
Methods/Technologies:perception, learning
Applications:service robotics



This project is vacant - apply now!