Software frameworks for artificial intelligence: comparsion of low-level and high-level approaches

Abstract

As nearly every artificial intelligence application is based on a framework, using the best fitting one for the task is key in developing an efficient solution quickly. Since there are two main types of frameworks, based on low and high abstraction level approaches, these two types will get compared and evaluated throughout this paper using Tensorflow and Keras as representatives. Key features of artificial intelligence frameworks for industrial applications are performance, expandability, abstraction level and therefore ease of use for rapid prototyping. All those features are major factors to keep development time and costs as low as possible, while maximizing product quality. To evaluate both approaches by these criteria a neural network classifying handwritten digits is implemented.