Implementing Neural Networks by Using the DataFlow Paradigm

• Published in: DataFlow Supercomputing Essentials pp. 3 - 44
• Main Authors: Milutinovic, Veljko, Kotlar, Milos, Stojanovic, Marko, Dundic, Igor, Trifunovic, Nemanja, Babovic, Zoran
• Format: Book Chapter
• Language: English
• Published: Cham Springer International Publishing 2017
• Series: Computer Communications and Networks
• Subjects: Control Flow Paradigm; Dataflow Implementation; Dataflow Paradigm; Execution Graph; FPGA Hardware
• ISBN: 9783319661247; 3319661248
• ISSN: 1617-7975; 2197-8433
• DOI: 10.1007/978-3-319-66125-4_1

In this chapter we will present one implementation of Neural Networks using dataflow paradigm. The dataflow paradigm presents a new approach to BigData applications. Existence of BigData is one of biggest application problems in many fields: financial engineering, geophysics, medical analysis, air flow simulations, data mining, and many others. Most of these applications are based on Neural Networks, and that being said, the way a network is implemented is crucial for the application performance. Such applications pay more attention to data than to the process itself. In order to be able to perform correct predictions, Neural Networks should be trained first. In some cases, they spend a lot of execution time on training process. The main challenge is finding a way to process such big quantities of data. Regardless of which level of parallelism is achieved, the execution process is essentially slow. In this chapter, the dataflow paradigm is presented as an alternative paradigm in solving this problem.