A fast learning algorithm for deep belief nets
Neural Computation
A Cortical Mechanism for Triggering Top-Down Facilitation in Visual Object Recognition
Journal of Cognitive Neuroscience
ICML '09 Proceedings of the 26th Annual International Conference on Machine Learning
Learning classifier systems: a complete introduction, review, and roadmap
Journal of Artificial Evolution and Applications
Learning Deep Architectures for AI
Foundations and Trends® in Machine Learning
Why Does Unsupervised Pre-training Help Deep Learning?
The Journal of Machine Learning Research
Proceedings of the 12th annual conference companion on Genetic and evolutionary computation
ICANN'10 Proceedings of the 20th international conference on Artificial neural networks: Part III
Visual Object Recognition
XCS-based versus UCS-based feature pattern classification system
Proceedings of the 14th annual conference on Genetic and evolutionary computation
Hi-index | 0.00 |
In order to allow more flexible and general learning, it is an advantage for artificial systems to be able to discover re-usable features that capture structure in the environment, known as Deep Learning. Techniques have been shown based on convolutional neural networks and stacked Restricted Boltzmann Machines, which are related to some degree with neural processes. An alternative approach using abstract representations, the ARCS Learning Classifier System, has been shown to build feature hierarchies based on reinforcement, providing a different perspective, however with limited classification performance compared to Artificial Neural Network systems. An Abstract Deep Network is presented that is based on ARCS for building the feature network, and introduces gradient descent to allow improved results on an image classification task. A number of implementations are examined, comparing the use of back-propagation at various depths of the system. The ADN system is able to produce classification error of 1.18% on the MNIST dataset, comparable with the most established general learning systems on this task. The system shows strong reliability in constructing features, and the abstract representation provides a good platform for studying further effects such as as top-down influences.