Synthesis and Analysis of Elementary Algorithms for a Differential Neural Computer
A. Kh. Rakhmatullin, R. F. Gibadullin
Abstract
Abstract Recurrent neural networks have found their application in a wide range of tasks related to the processing of sequential data sets, because of using of an internal state, which depends on both the current input data and the state at the previous iteration. DeepMind group proposed a new approach combining the attention mechanism and external memory. It became a big step in the development of this class of networks. The architecture called the Neural Turing Machine, as well as the later improved Differential Neural Computer (DNC) model, are an alternative to the classical Turing machine, with the exception that all its operations are differential. During all the opportunities that open, researchers faced with incomplete knowledge of its fundamental capabilities. In this thesis proposed a research method based on the use of DNC to solve basic elementary algorithms and analyzing the obtained characteristics of the model in comparison with classical algorithms for a Turing machine. The results obtained in this paper provide researchers with both practical advice on the use of DNC, shows weaknesses, and open new directions for the further improvement of this neural network architecture.