Exploring the Correlation Between Random Convolutional Architectures and the Trained Equivalent

In this paper we explore the correlation between Convolutional Neural Network (CNN) architectures with random weights in the convolutional layers to the same architectures with trained weights. We show that this correlation extends to deep CNN architectures of up to 10 or even 12 layers to the extent that untrained model accuracy could be a useful proxy for trained model accuracy. We also find that for models with fewer layers much of this relationship comes from the strong correlation between the number of features output from the final CNN layer and final accuracy. With 10 and 12 layers there is a moderate correlation even when the size of the fully connected layer is held constant. We anticipate our findings in extending these correlations to deeper networks will be useful in designing faster Neural Architecture Search (NAS) models. Analytically solving for the weights of the final prediction layer is orders of magnitude faster than training the weights via backpropagation.