Using Automatic Programming to Improve Neural Networks
Advances in Artificial Neural Networks (ANNs) have resulted in numerous new technologies in recent years. Many of these advances are the result of being bet- ter able to build functional neural network architectures using building blocks that have existed, sometimes, for decades. Only fairly recently have enhance- ments of the core building blocks, such as the new neuronal models Rectified Linear Unit (ReLU), and Exponential Linear Unit (ELU), become necessary to fully exploit both the potential of existing architectures and novel architectures. The slow progress in improving the core building blocks of neural networks is caused by the complex interactions in the network that make it difficult for humans to discover true improvements.
This project explores using the automatic programming system ADATE to automatically design improvements for neural network.