E.g. primitive operations like convolution and pooling.<p>From Neural Architecture Search: A Survey[0], first published in 2018:<p>"Moreover, common search spaces are also based on predefined building blocks, such as different kinds of convolutions and pooling, but do not allow identifying novel building blocks on this level; going beyond this limitation might substantially
increase the power of NAS."<p>I wonder now if somebody tried that since this was written or if anybody has some thoughts about the feasibility of this idea.<p>We've been using these methods for years. I doubt the first few breakthroughs are the most optimal possible.<p>[0]https://arxiv.org/abs/1808.05377