Theory and Practice of Natural Computing by Carlos Martín-Vide & Geoffrey Pond & Miguel A. Vega-Rodríguez

Author:Carlos Martín-Vide & Geoffrey Pond & Miguel A. Vega-Rodríguez
Language: eng
Format: epub
ISBN: 9783030345006
Publisher: Springer International Publishing

4 Experimental Setup

EA4CNN is implemented in C++. We utilized the parallel training of CNNs supported in TinyDNN (by means of OpenMP and SSE instructions). Experiments were executed on a computer node containing two Intel Xeon E5-2680v3 processors @ 2.5 GHz, 128 GB RAM and 24 threads. As the entire neuroevolution process is very time consuming (an average run in which 750 candidate CNNs are evaluated takes almost 72 h for CIFAR-10), we typically generated only 50 populations of 15 individuals and performed only five independent runs for a particular setup. Hence, most EA parameters and CNN (hyper)parameters were set up on the basis of preliminary results from several test runs.

EA4CNN was evaluated using MNIST (10 digit classes) and CIFAR-10 (10 image classes) classification problems. MNIST consists of pixel grayscale images of handwritten digits and includes 60 000 training images and 10 000 test images. In CIFAR-10, the numbers of training and test images are 50 000 and 10 000, respectively, and the size of images is pixels. For our purposes, these data sets were divided into three parts in such a way that there are 75 % vectors in , 10 % vectors in and and 15 % in .

The basic setup of EA parameters is as follows: , , , , , , for MNIST and 0.60 for CIFAR-10. Mutation operators MO1 – MO6 are used with the probabilities 0.41, 0.07, 0.03, 0.29, 0.10, and 0.10, respectively.

Table 1 summarizes the initial CNN hyperparameters for both data sets. Randomly generated networks of the initial populations contain from 1 to 8 layers in which all weights are randomly initialized to the close to zero values. TinyDNN utilizes the stochastic gradient descent learning method.Table 1.The initial setting of CNN hyperparameters in EA4CNN. The hyperparameters given in the first part of the table can be modified during the evolution.

Download

Copyright Disclaimer:
This site does not store any files on its server. We only index and link to content provided by other sites. Please contact the content providers to delete copyright contents if any and email us, we'll remove relevant links or contents immediately.