Deep learning has shown promising performance in addressing diverse real-world problems. The achievements of such algorithms are contributed greatly by their deep models. However, designing an optimal model often requires rich domain knowledge on both the investigated data and the general data analysis domain, which is not necessarily held by each interested end-user. In addition, the deep model defined for the task is often not reusable, and a new model must be redesigned for data with a slightly changed scenario. These facts collectively promote the development of techniques for automatically designing the models. In literature, this kind of automation is generally formulated as an optimization problem with non-convex and non-differentiable characteristics, challenging the design of proper optimization algorithms for solving the problem.

Evolutionary computation (EC) approaches, particularly genetic algorithms, particle swarm optimization and genetic programming, have shown superiority in addressing real-world optimization problems due largely to their powerful abilities in searching for global optima, dealing with non-convex/non-differentiable problems, and requiring no rich domain knowledge. EC approaches have great potential in automatically designing the optimal deep models. Indeed, there have been lots of EC works applied to the automation during past years, such as the architectures of deep neural networks. However, the development is still in the infant, and there still large room of designing novel efficient and effective EC methods to obtain satisfactory models.

The theme of this special session aims to bring together researchers investigating methods and applications in evolutionary deep learning. Particularly, the methods focus on effective and efficient representations, search mechanisms, optimization techniques, and fitness evaluation for automating the design of deep models. Authors are invited to submit their original work to this special session.