Book Review - Evolutionary Deep Learning
Book Review: Evolutionary Deep Learning by Micheal Lanham (Manning Publications , 2023)
A while ago, I read a book that was provided to me by Manning Publications (Thanks Adriana Sabo & Ana Romac ) titled “Evolutionary Deep Learning” by Micheal Lanham, under MEAP review.
I am writing this review based on my understanding on an upcoming field in AI. It is by no means an expert advice or a shout-out for sale :)
The book talks about implementing the concepts of Genetic Algorithms (GA) and Evolutionary Computation (EC) into Deep Learning(DL). The author has extensively touched the pain-points of architecture and training of Deep Learning models such as Optimizing loss functions, Hyper-parameter tuning and AutoML to name a few.
Key Takeaways
• **EC + DL = EDL**
Although, last decade has been fascinating for DL models, they have been in the limelight for some reason or the other, disrutped every indusrty vertical but practicioners of Data Science and ML developers very well knew the amount of task required to design and efficiently train such models. Optimizing DL models is driven by heuristic search at present which is a costly effort both in terms of compute cost and human expertise, so there is “a long way forward”. To overcome some issues, amalgamation of bio and nature inspired techniques with Deep Learning has been advocated in this book.
• Automated Hyperparameter Optimizations and Evolutionary Optimizations
EDL provides set of tools to automate hyperparameter search and architecture optimisation using GA , EC and PSO. Neuroevolution Optimization is an umbrella term for HPO, parameter optimization and network architecture optimization and selection.
• Applying EC to VAEs and Generative modeling for representation learning
Neuroevolution can be used to build a layered architecture that defines encoder and decoder sections in Variational AutoEncoders. In Generative Adversarial Networks, Genetic algorithms(GA) can be used to balance the training of a discriminator and generator and avoid mode collapse
The author has beautifully compared the traditional approaches with evolutionary ones to make the reader more comfortable in learning the applications of EC and GA to DL models. Although there is a little catch, the reader must have some familarity with Python, Visualisations using Python and some Data Science concepts along with some thoughts about Darwin’s Evolutionary Theory and how reproduction is carried out genetically in biological organisms using DNA, genes, and chromosomes.
Tip:A definite read for Data Scientists and ML developers who would like to explore/experiment new approaches for efficient, cost-effective and accurate Deep Learning models.
Book’s Link : https://www.manning.com/books/evolutionary-deep-learning