Novetta is using deep learning to automate advanced analytic pipelines for our customers. These new capabilities identify important pieces of information across large text, image, and video data sets, a challenge faced by many of our Defense and Intelligence Community customers.
We developed Argo, a standardized, Docker-based model training and deployment pipeline, to enable rapid evaluation and deployment of state-of-the-art deep learning models. Argo enables models to be deployed as endpoints and, when used with Amazon SageMaker, it forms the foundation for us to deploy, scale, and manage our deep learning models. Simplifying infrastructure setup through Docker and AWS allows our data scientists to focus on creating the best models possible without having to manage cloud resources.
Easy to Build, Maintain, and Deploy Advanced Deep Learning Models
Novetta’s Machine Learning Center of Excellence built a Natural Language Processing (NLP) capability to suggest customer-specific topic labels and tag quotes from news articles for our Novetta Mission Analytics open source intelligence platform. We developed a deep learning prototype that approximated human-level performance on this task. Using the Argo framework and Amazon SageMaker, we seamlessly transitioned from prototype to production deployment in just a day. Simplifying the maintenance of existing models and deployment of new models is essential to our customers, as we will eventually have dozens of models in production supporting Novetta Mission Analytics.
Framework for Quickly Building New Models
Working with the latest frameworks to install a deep learning library (e.g., Tensorflow, PyTorch, Caffe) can be complicated and error-prone. This only gets more difficult when algorithms and models are deployed across multiple systems and environments. Amazon SageMaker addresses some of these challenges by providing pre-built environments for many widely-used deep learning frameworks. When more custom development environments are needed, such as IoT devices for edge computing, we utilize Argo to deploy models in a host-agnostic fashion. The code we write to train and test models on our development machines is immediately and seamlessly compatible with Amazon SageMaker and Amazon EC2 instances. We are currently using the Argo pipeline with libraries and frameworks including Keras, PyTorch, and fast.ai.
Using Models in Production at Scale
Deployed models can be backed by CPU or GPU instances, and models can process data in batch mode or as a persistent endpoint. This flexibility supports both large, periodic data processing jobs and always-on, sub-second responses in mission-critical applications. Repeatable tasks traditionally performed by human analysis can now be automated, freeing up valuable resources to focus on more complex analytical challenges.