@article{Chung_Feiguth_Wong_2018, title={Mitigating Architectural Mismatch During the Evolutionary Synthesis of Deep Neural Networks}, volume={4}, url={https://openjournals.uwaterloo.ca/index.php/vsl/article/view/323}, abstractNote={<p>Evolutionary deep intelligence has recently shown great promise<br>for producing small, powerful deep neural network models via the<br>organic synthesis of increasingly efficient architectures over suc-<br>cessive generations. Existing evolutionary synthesis processes,<br>however, have allowed the mating of parent networks independent<br>of architectural alignment, resulting in a mismatch of network struc-<br>tures. We present a preliminary study into the effects of architec-<br>tural alignment during evolutionary synthesis using a gene tagging<br>system. Surprisingly, the network architectures synthesized using<br>the gene tagging approach resulted in slower decreases in perfor-<br>mance accuracy and storage size; however, the resultant networks<br>were comparable in size and performance accuracy to the non-<br>gene tagging networks. Furthermore, we speculate that there is a<br>noticeable decrease in network variability for networks synthesized<br>with gene tagging, indicating that enforcing a like-with-like mating<br>policy potentially restricts the exploration of the search space of<br>possible network architectures.</p>}, number={1}, journal={Journal of Computational Vision and Imaging Systems}, author={Chung, Audrey G. and Feiguth, Paul and Wong, Alexander}, year={2018}, month={Dec.}, pages={3} }