Deep Learning in Closed Loop
Lead Research Organisation:
University of Glasgow
Department Name: School of Engineering
Abstract
Deep learning, as a branch of machine learning has brought about many advances in the field of AI. It is a complex and multi-layered 'artificial neural network' that takes advantage of 'big data' resources and sufficient computing capacity for 'supervised' training using 'backpropagation algorithm'.
Deep learning aims at minimising the error generated through training iterations. Therefore, it is sensible to nest this algorithm within a closed loop feedback system where the supervised error determination in replaced by a PID controller that provides the error to be backpropagated; in turn, the PID controller is improved by the learning process.
This coupling of 'deep learning algorithm' and 'closed loop control system' can be tested on a simple robotic system where a path in followed by a moving robot. Nonetheless, this innovative approach has revolutionary potentials in the field of machine learning and AI.
Deep learning aims at minimising the error generated through training iterations. Therefore, it is sensible to nest this algorithm within a closed loop feedback system where the supervised error determination in replaced by a PID controller that provides the error to be backpropagated; in turn, the PID controller is improved by the learning process.
This coupling of 'deep learning algorithm' and 'closed loop control system' can be tested on a simple robotic system where a path in followed by a moving robot. Nonetheless, this innovative approach has revolutionary potentials in the field of machine learning and AI.
Organisations
| Description | We have established a direct implementation of deep learning into a closed loop system preserving its continuous processing for the first time. The performance of this learning paradigm is tested using a line-follower both in simulation and on a real robot that show very fast and continuous learning. |
| Exploitation Route | This learning paradigm can be sued in more complex robotic settings where the task is more involved for example a robotic arm for an stroke patient. Also the algorithm itself be taken a step further and be developed into one for discrete environment in which the learner makes hard (binary) decisions. |
| Sectors | Other |