Machine Learning - Research (EN)

NucleusResearch.com 4 Document Number: T147 October 2019 ▪ The number of people involved and their respective roles on project teams In total, Nucleus conducted in-depth interviews with 32 deep learning experts, many of whom were responsible for multiple projects concurrently, representing 316 unique projects. D E E P L E A R N I N G I N T H E C L O U D Production-scale deep learning workloads involve processing thousands or millions of example data to train the model. This is massively computationally expensive, especially for complex input data like images or video, and most organizations cannot afford to build and maintain high-performance computing systems with parallelized CPUs or GPUs to perform the calculations. As a result, organizations look to the cloud to access the resources and infrastructure they need. This year we found that 96 percent of deep learning projects are running in a cloud environment—this mirrors the finding last year, but with 177 projects in 2018 growing to 316 in 2019 it still demonstrates strong customer momentum to the cloud for deep learning. Of workloads that are in production on live data, 98 percent are in the cloud. For organizations that aren't fully in the cloud, a common deployment strategy involves developing a small-scale model on hardware on-site before migrating to the cloud for production. BENEFITS OF THE CLOUD With 96 percent of deep learning projects running in the cloud, clearly customers recognize the value of the approach. We asked respondents to identify the main benefits of running deep learning in the cloud. The responses clustered around three key themes: ▪ Cost savings from avoided hardware, personnel, and energy costs. This was the most common response, cited by every interviewee. Deep learning requires massive amounts of compute; building and maintaining hardware systems that can perform deep learning at scale requires dedicated IT professionals; physically running the hardware to train deep learning models consumes thousands of hours of CPU and GPU time—the electricity cost alone is often prohibitive. With the cloud, users pay for the resources they use without the associated costs. ▪ Ability to collaborate and work in distributed teams. Models deployed in the cloud can be accessible to all permissioned users, regardless of physical location. This speed up model development, especially across remote teams that are becoming increasingly common.

• Cover