Transfer Learning in GMDH-Type Neural Networks

Recognition of difficult patterns with the accuracy comparable to that of the human brain is a challenging problem. The ability of the human to excel at this task has motivated the use of Artificial Neural Networks (ANNs) which under certain conditions provide efficient solutions. ANNs are still unable to use the full potential of modular and holistic operations of biological neurons and their networks. The ability of neurons to transfer learned behaviour has inspired an idea to train ANN for a new task by using the behaviour patterns learnt from a related task. The useful patterns transferred from one task to another can significantly reduce the time needed to learn new patterns, and gives the neurons the ability to generalise instead of memorising patterns. In this paper we explore the ability of transfer learning for a face recognition problem by using Group Method of Data Handling (GMDH) type of Deep Neural Networks. In our experiments we show that the transfer learning of a GMDH-type neural network has reduced the training time by 31% on a face recognition benchmark.