[machine_learning_mastery系列]better_deep_learning

Modern open source libraries for developing neural network models are amazing. Gone are the days where we might spend weeks debugging the translation of poorly documented mathematics into code in the hopes of getting even the simplest model running. Today, using libraries like Keras, we can define and begin fitting a Multilayer Perc eptron, Convolutional or even Recurrent Neural Network model of arbitrary complexity in minutes. While defining and fitting models has become trivial, getting good performance with a neural network model on a specific problem remains challenging. Traditionally, configuring neural networks in order to get good performance was referred to as a dark art. This is because there are no clear rules on how to best prepare data and configure a model for a given problem. Instead, experience must be developed over time from working on many different projects. Nevertheless, neural networks have been used in academia and industry for decades now and there are a suite of standard techniques, tips, and configurations that you can use to greatly increase the likelihood of getting better-than-average performance with a neural network model. I wrote this book to pull together the best classical and modern techniques in order to provide a playbook that you can use to get better performance on your next project using deep learning neural networks. A lot has changed in the last 5 to 10 years and there are activation functions, regularization methods, and even new ensemble methods that result in remarkably faster learning, lower generalization error, and more robust results. I hope that you’re as excited as me about the journey ahead. Jason Brownlee 2018 eptron, Convolutional or even Recurrent Neural Network model of arbitrary complexity in minutes. While defining and fitting models has become trivial, getting good performance with a neural network model on a specific problem remains challenging. Traditionally, configuring neural networks in order to get good performance was referred to as a dark art. This is because there are no clear rules on how to best prepare data and configure a model for a given problem. Instead, experience must be developed over time from working on many different projects. Nevertheless, neural networks have been used in academia and industry for decades now and there are a suite of standard techniques, tips, and configurations that you can use to greatly increase the likelihood of getting better-than-average performance with a neural network model. I wrote this book to pull together the best classical and modern techniques in order to provide a playbook that you can use to get better performance on your next project using deep learning neural networks. A lot has changed in the last 5 to 10 years and there are activation functions, regularization methods, and even new ensemble methods that result in remarkably faster learning, lower generalization error, and more robust results. I hope that you’re as excited as me about the journey ahead. Jason Brownlee 2018