A view on automated neural graph topology generation and a viable direction of innovation
Andrei Ionut Damian, Laurentiu Piciu, Nicolae Ţăpuş · 2020
Training deep neural networks requires knowledge and ample experimental time, as well as computational resources due to the search process of the architectural design. In this work, we review the current main research directions in the area of network architecture and finally propose in contrast a novel architecture, namely MultiGatedUnit, that uses directly learnable self-gating mechanisms for automated graph topology generation, currently in research and experimentation phase.