[Project in collaboration with Norwegian OpenAI Lab]
Deep Learning has achieved incredible results. However, the success of DL depends on finding the right architecture for the task at hand. One example is Google Inception v.3, which is made of many handcrafted modules precisely engineered and connected. In addition, DL targets more and more complex problems and therefore more complex architectures are needed. Hybrid architectures made of deep layers trained through gradient descent and networks evolved through neuroevolutionary techniques have been proposed [1, 2]. In addition, the evolution of deep learning architectures have been shown successful . Modular architectures have also been utilized to create deep networks able to solve and adapt to different tasks .
In this project, we want to combine the benefits of deep learning and neuroevolution, by creating an incrementally complex evolved deep learning architecture that can learn to solve different tasks, i.e. meta-learning  (learning to learn). Meta-learning has recently emerged as a major research direction towards artificial general intelligence (AGI).
This project is in collaboration with Ass.Professor Stefano Nichele @ HiOA
 DLNE: A Hybridization of Deep Learning and Neuroevolution for Visual Control
 Evolving deep unsupervised convolutional networks for vision-based reinforcement learning
 Evolving Deep Neural Network
 PathNet: Evolution Channels Gradient Descent in Super Neural Networks