Q-NAS Revisited: Exploring Evolution Fitness to Improve Efficiency

A neural network architecture optimizer based on DARTS and generative adversarial learning

“…Later, a new phase of research on NAS algorithms was launched. During this research period, most of the works focused on evolutionary algorithms [11][12][13]. In this period, the universal and representative network layers were relatively shallow, the neurons in each layer were small in number, and the overall network architectures were simple chain structures.…”

Quantum-Inspired Neural Architecture Search

“…Every generation, we create C(t) with size K, then we keep the K best individuals from [C(t) ∪ P (t)]. Our studies in [72] analyzed the elitism selection, which did not show any improvement over this steady-state method. Furthermore, the variation that elitism promoted in the population only brought noise and no additional benefit.…”

“…In this section, we complete our analysis of Q-NAS applied to CIFAR-10 by studying the addition of an early-stopping mechanism. Our previous work [72] showed that Q-NAS could benefit from a simple early-stopping method, significantly reducing the total runtime, while maintaining the level of test accuracy. These results, however, were based on Q-NAS runs that took no more than 60 hours to complete.…”

“…In the previous analysis, we related these plateaus not to stagnation but to the difficulty in finding architectures with a stand-out performance. The early-stopping criterion considers the small increments in fitness [72]: stop if the best individual fitness does not improve above a threshold of 0.005 for 80 generations.…”