Page 13 - Artificial Intelligence in the Age of Neural Networks and Brain Computing

Introduction xxv

In Chapter 15, Miikkulainen et al. present a novel automated method for
designing deep neural network architecture. The main idea is based on neuroevolu-
tion to evolve the neural network topology and parameters. In the proposed work,
neuroevolution is extended to evolve topology, components (modules), and hyper-
parameters. The method is applied to both feedforward architectures like CNN
and also to recurrent neural networks (with LSTM units). The proposed method is
tested on standard image tasks (object recognition) and natural language tasks
(image captioning), demonstrating comparable results as state-of-the-art methods.
This chapter provides a great overview of evolutionary methods developed at
Sentient technologies for the design and optimization of deep neural networks.