Search Tree Pruning for Progressive Neural Architecture Search (Student Abstract)

This transition graph defines valid neural network layer sequences in the search tree. Output is a 10-neuron, Dense, softmax layer. Experiments After four days of running on an Intel i7 8th generation CPU, 61 different network architectures were created with the best having an accuracy of 91.9% on the Fashion-MNIST dataset. Figure 4 compares our algorithm's best accuracy to other networks trained on the same dataset. Our model has <2% difference in accuracy compared to GoogleNet, the best non-preprocessed model with an accuracy of 93.7% [1] . Finally, our algorithm produced a network with less parameters in less time on a CPU compared to other benchmark models-VGG16, GoogleNet, and AlexNet. This can be seen in Table 1 .