EEEA-Net: An Early Exit Evolutionary Neural Architecture Search

The goals of this research were to search for Convolutional Neural Network (CNN) architectures, suitable an on-device processor with limited computing resources, performing at substantially lower Architecture Search (NAS) costs. A new algorithm entitled Early Exit Population Initialisation (EE-PI) Evolutionary Algorithm (EA) was developed achieve both goals. EE-PI reduces the total number parameters in process by filtering models fewer than maximum threshold. It will look a model replace those more Thereby, reducing parameters, memory usage storage and processing time while maintaining same performance or accuracy. reduced 0.52 GPU day. This is huge significant achievement compared NAS 4 days achieved using NSGA-Net, 3,150 AmoebaNet model, 2,000 NASNet model. As well, networks (EEEA-Nets) yield network architectures minimal error computational cost given dataset as class algorithms. Using EEEA-Net on CIFAR-10, CIFAR-100, ImageNet datasets, our experiments showed that lowest rate among state-of-the-art models, 2.46% 15.02% 23.8% dataset. Further, we implemented image recognition architecture other tasks, such object detection, semantic segmentation, keypoint detection and, experiments, EEEA-Net-C2 outperformed MobileNet-V3 all these various tasks. (The code available https://github.com/chakkritte/EEEA-Net).