Automated design of the deep neural network pipeline

Loading...
Thumbnail Image

Date

Authors

Journal Title

Journal ISSN

Volume Title

Publisher

University of Pretoria

Abstract

Deep neural networks have been shown to be very effective for image processing and text processing. However the big challenge is designing the deep neural network pipeline, as it is time consuming and requires machine learning expertise. More and more non-experts are using deep neural networks in their day-to-day lives, but do not have the expertise to parameter tune and construct optimal deep neural network pipelines. AutoML has mainly focused on neural architecture design and parameter tuning, but little attention has been given to optimal design of the deep neural network pipeline and all of its constituent parts. In this work a single point hyper heuristic (SPHH) was used to automate iii the design of the deep neural network pipeline. The SPHH constructed a deep neural network pipeline design by selecting techniques to use at the various stages of the pipeline, namely: the preprocessing stage, the feature engineering stage, the augmentation stage as well as selecting a deep neural network architecture and relevant hyper-parameters. This work also investigated transfer learning by using a design that was created for one dataset as a starting point for the design process for a different dataset and the effect thereof was evaluated. The reusability of the designs themselves were also tested. The SPHH designed pipelines for both the image processing and text processing domain. The image processing domain covered maize disease detection and oral lesion detection specifically and text processing used sentiment analysis and spam detection, with multiple datasets being used for all the aforementioned tasks. The pipeline designs created by means of automated design were compared to manually derived pipelines from the literature for the given datasets. This research showed that automated design of a deep neural network pipeline using a single point hyper-heuristic is effective. Deep neural network pipelines designed by the SPHH are either better than or just as good as manually derived pipeline designs in terms of performance and application time. The results showed that the pipeline designs created by the SPHH are not reusable as they do not provide comparable performance to the results achieved when specifically creating a design for a dataset. Transfer learning using the designed pipelines is found to produce results comparable to or better than the results achieved when using the SPHH without transfer learning. Transfer learning is only effective when the correct target and source are chosen, for some target datasets negative transfer occurs when using certain datasets as the transfer learning source. Future work will include applying the automated design approach to more domains and making designs reusable. The transfer learning process will also be automated in future work to ensure positive transfer occurs. The last recommendation for future work is to construct a pipeline for unsupervised deep neural network techniques instead of supervised deep neural network techniques.

Description

Dissertation (MSc (Computer Science))--University of Pretoria, 2021.

Keywords

Automated design, Transfer learning, Deep neural network pipeline, Text classification, Image segmentation

Sustainable Development Goals

Citation

*