<p>In this paper we investigate the impact that altering the convolutional layer topology has upon the performance of computer vision tasks using a variety of widely used benchmark image datasets. Despite the widespread convention in convolutional neural networks, of incrementally doubling the filter count at each layer, there is little evidence substantiating the superiority of this method over other possible topologies. Our research reveals that a contrarian strategy—reducing the filters by half—can achieve performance on par with, if not superior to, this usual approach. We have extended our investigation to include a variety of novel topological structures. These empirical results challenge the prevailing assumption, that the sequential doubling of number of filters in the network configuration will always yield the best results with all datasets. Our findings advocate for a more nuanced approach to neural network design, incorporating a flexible approach to filter topologies into workflows. This could potentially have a significant impact upon the architectural standards in deep learning for visual recognition tasks.</p>