Hitoshi Iba: Evolutionary approach to machine learning and deep neural networks: neuro-evolution and gene regulatory networks - Springer, 2018, Hardcover, 245 pp, ISBN: 978-981-13-0199-5

Petra Vidnerová

Introduction

The intersection of Evolutionary Computation (EC) and Machine Learning (ML) has become a focal point for researchers seeking to automate the design of complex artificial intelligence systems. Hitoshi Iba’s book, Evolutionary approach to machine learning and deep neural networks: neuro-evolution and gene regulatory networks (Springer, 2018), attempts to synthesize these two historically distinct fields. This review evaluates the book's success in bridging the gap between biological inspiration and modern deep learning architectures.

The Bidirectional Synergy of EAs and ML

The central premise of Iba’s work is the "synergy problem": the reciprocal relationship where Evolutionary Algorithms (EAs) and ML techniques enhance one another. This is framed as a two-way street:

  1. EAs for ML: Utilizing evolutionary processes to optimize neural network architectures and hyperparameters, a field known as neuro-evolution.
  2. ML for EAs: Leveraging machine learning as a meta-heuristic to improve the efficiency and search capabilities of evolutionary algorithms.

The book provides a roadmap for this intersection, which is increasingly relevant as the demand for automated machine learning (AutoML) grows.

Key Methodological Pillars

The book is structured around several core themes that highlight the diversity of hybrid approaches:

Neuro-evolution and Deep Architectures

Iba discusses foundational algorithms such as NeuroEvolution of Augmenting Topologies (NEAT) and its extension, hyperNEAT. The text also touches upon more contemporary developments like DeepNEAT, which aims to optimize the complex topologies required for deep learning.

Algebraic Genetic Programming

A notable and somewhat unexpected inclusion is the deep dive into using Gröbner bases for an algebraic approach to Genetic Programming (GP). This section moves beyond traditional "trial and error" evolution, suggesting a push toward more mathematically rigorous foundations for GP.

Bio-inspired Modeling and Robotics

The application of Gene Regulatory Networks (GRNs) to fields like bioinformatics and humanoid robotics is explored. This suggests that the future of AI may involve mimicking not only the brain's neurons but also the underlying chemical and genetic systems that regulate biological growth and adaptation.

Critical Assessment

The reviewer, Petra Vidnerová, highlights several strengths and weaknesses of the text:

Strengths: * Visual Communication: The book’s greatest asset is its heavy reliance on diagrams and illustrations, making complex algorithmic flows accessible to visual learners. * Chapter 4 (ML for EAs): This section is cited as the book's strongest, offering comprehensive information, up-to-date references, and well-documented performance results. * Accessibility: The text is largely self-explanatory, making it an excellent entry point for beginners or a useful reference for experts tracking hybrid trends.

Weaknesses: * Dated Content: Some discussions, particularly regarding NEAT and hyperNEAT, focus on algorithms from the early 2000s, which may feel outdated to researchers in the fast-moving deep learning space. * Uneven Depth: While niche topics like Gröbner bases receive significant attention, core deep learning concepts are sometimes "squeezed into hardly one page," which may disappoint specialists. * Reference Gaps: The reviewer notes the absence of foundational references in introductory chapters, such as John Koza’s seminal work on Genetic Programming. * Lack of Formality: Certain sections, such as those on GRNs, lack formal definitions at the outset, potentially hindering readers seeking mathematical precision.

Conclusion

Hitoshi Iba’s book serves as a solid high-level overview of the hybridization of evolutionary computation and machine learning. While it may lack the technical depth required by deep learning specialists or the formal rigor expected in specific mathematical sub-fields, its visual clarity and broad scope make it a valuable resource for those entering the field. For a complete understanding, readers are encouraged to supplement this text with standard machine learning literature to fill in the technical gaps regarding modern deep neural networks.

Summary updated

BibTeX
@article{vidnerova2019hitoshi,
  author = {Petra Vidnerová},
  title = {{Hitoshi Iba: Evolutionary approach to machine learning and deep neural networks: neuro-evolution and gene regulatory networks - Springer, 2018, Hardcover, 245 pp, ISBN: 978-981-13-0199-5}},
  journal = {Genet. Program. Evolvable Mach.},
  year = {2019},
  doi = {10.1007/s10710-019-09350-8},
}
Figure