Machine Learning - Research (EN)

NucleusResearch.com 3 Document Number: T147 October 2019 T H E S I T U A T I O N Machine learning encompasses the technology where a computer analyzes data to "learn" from experience without human involvement. Deep learning is a subset of machine learning; in machine learning, the computer is given data with a designated set of features to analyze, whereas in deep learning the computer is presented with unstructured data like text, audio, or video, and it determines by itself which features are relevant to the analysis. Put simply, the computer is provided with pairs of sample inputs and the corresponding outputs and is able to work backwards to find what operations are necessary to transform the input to the output. Modern deep learning models require massive amounts of compute and storage, making it prohibitive for most organizations to build these systems themselves. Thus, as we found throughout the course of this research, companies overwhelmingly look to leverage the cloud for deep learning projects. This approach allows businesses to buy the amount of data storage and compute power they need for the projects without having to purchase, configure, and maintain the infrastructure internally, producing significant cost savings over time. With a diverse and quickly growing user base due to the hype and potential of AI, the technology landscape changes quickly. New tools and methodologies are constantly becoming available. Therefore, enabling users to develop on the platform with maximum flexibility is key to long term success in the cloud-based machine learning market. Simply put, cloud platforms need to support the myriad tools and development frameworks that are in use today and tomorrow, with the requisite security and availability to adhere to data handling and privacy regulations. This is the third consecutive year of conducting this study, and over these three years we have seen transformative changes in model capabilities, compute power, and developer tools enabling new and exciting results. In the first year it was difficult to find organizations that had moved beyond preliminary development and proof-of-concept projects with deep learning. In 2018 we saw strong progress to this end with 14 percent of projects being classified as in production, handling live data. This year brought another leap forward to this effect with organizations ranging from 20-person startup to Fortune 100 global enterprises deploying deep learning to production, with 38 percent of percent of projects in production. Other aspects discussed in the interviews include: ▪ The goals and motivations behind the deep learning projects ▪ The deployment strategy and associated benefits ▪ The development frameworks, methods, and other tools being used ▪ The relative strengths and weaknesses of different models and frameworks

• Cover