A hybrid training method of convolution neural networks using adaptive cooperative particle swarm optimiser.

For solving the problem that it is easy to fall into the local minimum in Convolution Neural Networks (CNN) training, a hybrid training algorithm based on heuristic algorithm is proposed. Firstly, an Adaptive Cooperative Particle Swarm Optimisation (ACPSO) is proposed, which uses a learning automata to adaptively divide the subpopulation of the Cooperative Particle Swarm Optimisation (CPSO), and makes the decision variables with strong coupling relationship enter the same subpopulation. Then, the connection weights of CNN are considered as elements in particles and the CNN is trained by ACPSO algorithm. The output of the ACPSO algorithm is applied as the initial weight of the BP algorithm for the purpose of speeding up the training speed of the CNN. The experimental results show that the ACPSO-BP algorithm has achieved good results, and the recognition rate of the CNN is improved. Thus it has the potential to be applied to other deep learning fields.