Detail of publication

Abstract

The unreliable behaviour of deep neural networks for specific use cases slows down their deployment in real world applications. The mainstream way to catch up the remaining percents in performance relies on iterative tuning of hyper-parameters, collecting new data and increasing computational power, while the behaviour inside the trained model is usually kept shrouded in mystery. As the commonly used network architectures have been found un- necessarily complicated as well as limited, e.g. by the strict arrangment of neurons into layers, this work proposes an alternative method to work with neural principles and to design network architectures. The neural princi- ples are combined with the theory of multi-agent systems and reinforcement learning with the goal of generating tailored networks for given classification tasks. The tailored networks are believed to shake off redundant parts and to allow less restricted way of connecting neurons in comparison to standard structures. This could possibly improve the performance as well as enable targeted fixes in the network even after the training. The main loop of the algorithm has been already implemented and tested on two basic 2D exper- iments with expected results. The work is currently at the state of prooving the scalability to multidimensional classification tasks.