We propose a novel method to evolve autoencoding structures through genetic programming (GP) for representation learning on high dimensional data. It involves a partitioning scheme of high dimensional input representations for distributed processing as well as an on-line form of learning that allows GP to efficiently process training datasets composed of hundreds or thousands of samples. The use of this on-line learning approach has important consequences in computational cost given different evolutionary population dynamics, namely steady state evolution and generational replacement. We perform a complete experimental study to compare the evolution of autoencoders (AEs) under different population dynamics and genetic operators useful to evolve GP based AEs’ individuals. Also, we compare the performance of GP based AEs with another representation learning method. Competitive results have been achieved through the proposed method. To the best of the authors’ knowledge, this research work is a precursor within the field of evolutionary deep learning.

中文翻译：

---

通过遗传编程进化自动编码结构

我们提出了一种通过遗传编程（GP）进化自动编码结构的新方法，用于高维数据的表征学习。它涉及用于分布式处理的高维输入表示的分区方案以及允许 GP 有效处理由数百或数千个样本组成的训练数据集的在线学习形式。鉴于不同的进化种群动态，即稳态进化和世代更替，这种在线学习方法的使用对计算成本具有重要影响。我们进行了一项完整的实验研究，以比较不同种群动态和遗传算子下自动编码器 (AE) 的进化，这有助于进化基于 GP 的 AE 个体。还，我们将基于 GP 的 AE 的性能与另一种表示学习方法进行比较。通过所提出的方法已经取得了有竞争力的结果。据作者所知，这项研究工作是进化深度学习领域的先驱。