Abstract
Convolutional neural networks (CNNs) have shown great success in a variety of real-world applications and the outstanding performance of the state-of-the-art CNNs is primarily driven by the elaborate architecture. Evolutionary convolutional neural network (ECNN) is a promising approach to design the optimal CNN architecture automatically. Nevertheless, most of the existing ECNN methods only focus on improving the performance of the discovered CNN architectures without considering the relevance between different classification tasks. Transfer learning is a human-like learning approach and has been introduced to solve complex problems in the domain of evolutionary algorithms (EAs). In this paper, an effective ECNN optimization method with cross-tasks transfer strategy (CTS) is proposed to facilitate the evolution process. The proposed method is then evaluated on benchmark image classification datasets as a case study. The experimental results show that the proposed method can not only speed up the evolutionary process significantly but also achieve competitive classification accuracy. To be specific, our proposed method can reach the same accuracy at least 40 iterations early and an improvement of accuracy for 0.88% and 3.12% on MNIST-FASHION and CIFAR10 datasets compared with ECNN, respectively.
Highlights
Machine learning has shown great success in various real-world applications
At the early generations, our method can increase the optimization speed significantly, which is important when the learning time or computing resource is limited; (2) Within the case study of image classification tasks, it is demonstrated that the proposed cross-tasks transfer strategy (CTS)-Evolutionary convolutional neural network (ECNN) can obtain better results than the ECNN that starts from scratch and some manually-designed state-of-the-art methods do; (3) In the framework of the proposed CTS-ECNN, when a new task is encountered, we can extract knowledge from the optimized tasks
The proposed CTS-ECNN method, which is based on transfer learning, is evaluated using two benchmark image classification datasets
Summary
Machine learning has shown great success in various real-world applications. Convolutional neural networks (CNNs), which show overwhelmingly superiority among machine learning approaches, have been widely used in various real-world applications, such as image processing [1], engineering [2], health care [3,4], and cognitive science [5], etc. Convolutional neural network commonly consists of convolution, pooling, and fully-connected layers and are trained on the source dataset and applied to the target dataset. As is well-known, the success of CNNs mainly benefit from the improvement on fundamental CNN architectures, such as increasing the depth of neural networks, the employment of skip layers, and adding inner network structures, etc. The state-of-the-art CNN architectures with high performance are manually devised by experienced experts with trial-and-error. As designing efficient CNN architectures is a challenging process, researchers have developed algorithms to design the CNN architectures automatically, which aims to enhance the applicability and universality of CNNs
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a callDisclaimer: All third-party content on this website/platform is and will remain the property of their respective owners and is provided on "as is" basis without any warranties, express or implied. Use of third-party content does not indicate any affiliation, sponsorship with or endorsement by them. Any references to third-party content is to identify the corresponding services and shall be considered fair use under The CopyrightLaw.
