Machine Learning - Research (EN)

NucleusResearch.com 7 Document Number: T147 October 2019 performance it offers, even at petabyte-scale, and the flexibility of supported machines and its work to optimize all frameworks to perform well. RELATIONSHIP WITH AMAZON Interviewed customers also referenced their relationship with AWS as a driving factor behind the business decision. Many customers were already using AWS in other areas of the business and grew their investment to include deep learning. Deep learning requires a lot of data, and organizations already trust AWS with their data. Amazon also provides customers with the technology and best practices to complete their projects without locking them into any particular solution. Customers were quick to emphasize the flexibility of the platform— they can use the frameworks and libraries they find most comfortable without worrying that they are incompatible with AWS technology. One customer referenced the fact that Amazon is interested in studies like this as a demonstration of its commitment to maximizing customer value. Other users said: • "We built our entire business on AWS, so it would take a lot to motivate us to choose another provider. Our AWS [architecture] has grown with us, and the support has been solid every step of the way. We really feel like they're invested in our success which is key to any long-term partnership." • "With Amazon, we have access to customer use cases and industry expertise for our machine learning projects. The support team was instrumental to our efforts—they were able to answer all of our questions about frameworks, model choice, and infrastructure requirements." User profile – Digital Media and Mobile Game Company A digital media company that designs and sells mobile games uses deep learning to balance game difficulty with the projected revenue generated. It does this by using deep reinforcement learning to train a bot to play each game. Then it can monitor the bot's performance to estimate the difficulty of each level which it then uses to predict how many users will quit playing at each level. Of course, when a user stops playing, they stop generating ad revenue for the game vendor, so the company uses this process to balance game difficulty with projected revenue. AWS has been the company's cloud provider since 2008, so it was a no-brainer to use AWS for its deep learning. Both projects are built on PyTorch—PyTorch is the industry standard for deep reinforcement learning, and much of the current research is published with an accompanying PyTorch implementation.

• Cover