Evolutionary Neural Architecture Search

Deep-learning architectures always felt quite arbitrary to me. Not only the parameters, but also which layers to use and how they are combined. Classic hyperparameter search methods like grid search can in principle include the architecture itself, but then the search space simply becomes too big.

So this work took a new approach: an evolutionary architecture search that also decides what kind of deep-learning layer to use. Sequential architectures for the classification of time-series data are found, evaluated, and improved using evolutionary techniques. Compared to standard architectures in that field, the evolutionary evolved architecture achieves an accuracy improvement of up to 5.82 percent.

The full thesis is embedded below.