Evolving Character-Level DenseNet Architectures Using Genetic Programming

Densely Connected Convolutional Networks (DenseNet) have demonstrated impressive performance on image classification tasks, but limited research has been conducted using character-level DenseNet (char-DenseNet) architectures for text tasks. It is not clear what are optimal The iterative task of designing, training and testing char-DenseNets a time consuming that requires expert domain knowledge. Evolutionary deep learning (EDL) used to automatically design CNN the domain, thereby mitigating need This study demonstrates first work EDL evolve char-DenseNet A novel genetic programming-based algorithm (GP-Dense) coupled with an indirect-encoding scheme, facilitates evolution performant architectures. evaluated two popular datasets, best-evolved models benchmarked against four current state-of-the-art models. Results indicate evolves both datasets outperform in terms model accuracy three parameter size.