دانلود کتاب طراحی سیستم های یادگیری عمیق

A little more than a decade ago, we had the privilege of building some early end user–facing product features that were powered by artificial intelligence. It was a huge undertaking. Collecting and organizing data that would be fit for model training was not a usual practice at that time. Few machine learning algorithms were packaged as ready-to-use libraries. Performing experiments required running management manually and building out custom workflows and visualizations. Custom servers were made to serve each type of model. Outside of resource-intensive tech companies, almost every single new AI-powered product feature was built from scratch. It was a far-reaching dream that intelligent applications would one day become a commodity. After working with a few AI applications, we realized that we had been repeating a similar ritual each time, and it seemed to us that it made more sense to design a systematic way, with prototyping, for delivering AI product features to production. The fruit of this effort was PredictionIO, an open source suite of framework software that put together state-of-the-art software components for data collection and retrieval, model training, and model serving. Fully customizable through its APIs and deployable as services with just a few commands, it helped shorten the time required at every stage, from running data science experiments to training and deploying production-ready models. We were thrilled to learn that developers around the world were able to use PredictionIO to make their own AI-powered applications, resulting in some amazing boosts to their businesses. PredictionIO was later acquired by Salesforce to tackle a similar problem on an even larger scale. By the time we decided to write this book, the industry was thriving with a healthy AI software ecosystem. Many algorithms and tools have become available to tackle different use cases. Some cloud vendors such as Amazon, Google, and Microsoft even provide complete, hosted systems that make it possible for teams to collaborate on experimentation, prototyping, and production deployments at one centralized location. No matter what your goal is, you now have many choices and numerous ways to put them together. Still, as we work with teams to deliver deep learning–powered product features, there have been some recurring questions. Why is our deep learning system designed the way it is? Is this the best design for other specific use cases? We noticed that junior software engineers were the ones most often asking these questions, and we interviewed a few of them to find out why. They revealed that their conventional software engineering training did not prepare them to work effectively with deep learning systems. And when they looked for learning resources, they found only scant and scattered information on specific system components, with hardly any resources discussing the fundamentals of the software components, why they were put together the way they were, and how they worked together to form a complete system.