Semantics in Deep Neural-Network Computing

Abstract

Artificial Intelligence development is stepping into a new era due to the recent exciting achievements from neural network and statistical machine learning research communities. Statistic neural-computing based machine learning has been deemed as one of promising roads towards realizing the ideal of Artificial Intelligence promoted since last century. Learning is the key in making progress. Statistic machine learning is to obtain a probability distribution or a function from a set of training samples according to a certain optimization target over the training cost based on a predefined model. While there are many significant improvements in image, sound and text recognition and analyzing using neural network based learning strategies, a new open question emerges, that is, what is the next? To ask this question is to mean that the neural network solution is not an ultimate solution, and there will be more challenges to meet in coming future. We discussed these aspects of deep neural network research work in this paper and focused on semantics in deep neural-network computing models. We try to browse how semantics or knowledge are to be involved in deep neural network models and how the semantics and knowledge will be a key factor towards making more intelligent machines. We argue that priori semantics and knowledge in modelling a neural network is important, which could be the key for researchers to design intelligent machine models to perform complex tasks.