Machine Learning on AWS

MACHINE LEARNING ON AWS

Solving data-intensive text, image, and sensor-based challenges by deploying state-of-the-art machine learning techniques in AWS.

Built Using the Power of Amazon Web Services

Machine learning is having a transformative impact on advanced analytics, delivering breakthroughs in applications ranging from face recognition to predictive modeling. Harnessing the power of machine learning in real-world applications requires experience with complex, diverse datasets, whether for training, model development, or performance optimization.

We were among the first in the national security space to achieve AWS Machine Learning Competency status. This competency is a formal recognition of the quality of the machine learning solutions we develop for customers as well as our knowledge of the AWS ecosystem. AWS supports our rapid innovation efforts, whether prototyping on small CPU instances or scaling up to multi-GPU instances. Amazon SageMaker streamlines this process, allowing our machine learning researchers to focus on challenging problems instead of DevOps. When it’s time to deploy solutions, Amazon SageMaker endpoints simplify progression from testing and evaluation to production.

Our solutions extract meaningful content from structured and unstructured data, empowering analysts to deliver intelligence and insights. We address complex challenges using deep learning algorithms implemented with frameworks such as PyTorch and Tensorflow. Our machine learning engineers identify opportunities where machine learning tools can complement or replace conventional, rules-based, or analyst-dependent functions to increase the scale, speed, and accuracy of large-scale data analytics. We also conduct rigorous benchmarking and performance evaluations to identify machine learning approaches, platforms, and tools that are best-suited to address customer needs.

State of the Art Research

With new methods and algorithms released every day, it can be difficult to identify those capable of impacting customer missions. Our machine learning researchers stay abreast of the latest advances in natural language processing (NLP), computer vision, geospatial imaging, biometrics, and cyber through independent research, collaborative groups, and formal seminars. Through those channels, researchers prioritize advances for additional validation against our customers’ mission sets.

Rapid Prototyping

Our customers need to benefit from new ML methods and algorithms as soon as possible. We create functional machine learning prototypes in days, demonstrating immediate benefits and serving as the foundation for customer-supporting teams to deliver final solutions.

State of the Art Research

Rapid Prototyping

AWS-Based Solutioning

As an AWS Advanced Tier Consulting Partner, we deliver cloud-based machine learning solutions using Amazon SageMaker. By offering on-demand access to a wide range of algorithms and frameworks, SageMaker simplifies implementation of machine learning for many of our customers. We also use a variety of AWS services including EC2, S3, EFS, Lambda, and ML services including Rekognition, Comprehend, and Transcribe.

Analytics at the Edge

While the cloud provides near-infinite computing resources at the push of a button, our customers often operate in austere environments without ready access to the internet or cloud. Those serving in disconnected environments often need the same ability to analyze data as their counterparts at home. Novetta is working closely with AWS to pioneer the ability to put machine learning services in the hands of our customers, no matter where they are.

AWS-Based Solutioning

Analytics at the Edge

Building on our extensive machine learning practice spanning dozens of customers and mission sets, the Machine Learning Center of Excellence (ML COE) is a focal point for engineering, research and development, and collaboration in machine learning. Staffed by data scientists and software engineers building advanced machine learning prototypes and solutions, the ML COE provides our workforce with training and mentorship in machine learning to build and maintain skills essential to customer success. Members of the ML COE hold AWS solutions architect, developer, big data, and machine learning speciality certifications demonstrating mastery of the platform and the ability to harness the capabilities of AWS. These efforts have resulted in Novetta being awarded the first AI/ML Public Sector Consulting Partner Award and being recognized as experts in the field through numerous conference speaking engagements.

Innovation Lab

Novetta uses the Center to conduct targeted research into topics such as transfer learning for Natural Language Processing (NLP), computer vision (CV), cyber, audio, and biometrics. Products of the Center’s R&D efforts are made available to government and industry partners through a hands-on Innovation Lab. Government and industry partners can explore the latest in machine learning technologies through interactive demonstrations customized to highlight the particular interests of the customer.

The Center regularly hosts ML Lunch and Learns, publishes blog posts, shares notes from conferences and university speaking events, and posts a monthly newsletter available to anyone interested in learning more about the field.

Increase the Scale, Speed, and Accuracy of your Analytics

Automating Unstructured Text Classification

New NLP algorithms have significantly improved upon the prior state of the art for classifying unstructured text (e.g., news articles, emails, reports). Within days of their release, Novetta’s machine learning researchers were using Amazon SageMaker to evaluate how these algorithms could benefit our customers and product teams.

After conducting an evaluation on internal data, our machine learning engineers worked closely with our Novetta Mission Analytics team to integrate high-performing models into their production application hosted on AWS. Models were easily transitioned to their production environment using SageMaker endpoints. The solution increased the efficiency, accuracy, and quality of analysts’ hand labeling of news articles. Further refinements, including a method for incorporating structured data into the model, increasing accuracy by 20%.

The teams also worked together to create an automated training pipeline that significantly decreased the level of effort it takes to train and deploy new models. This improved pipeline has made it easy to develop solutions for other use cases, from identifying terrorist propaganda to classifying company descriptions by North American Industry Classification System (NAICS) code.

Read the Blog: NLP Transfer Learning on SageMaker

Applying Graph Analytics to Make Sense of Organizational Data

Relationships are a natural part of data encountered in most organizations, but that information has traditionally been challenging to exploit. In 2019 researchers released new algorithms that work better on graph data, leading to significant improvements in accuracy on tasks such as entity classification and relationship prediction.

Researchers from the ML COE worked closely with a customer team to analyze company and key personnel to better identify high-risk entities that had not been previously flagged in the database. Using entity data resolved by Novetta Entity Analytics, the team trained a model to learn which relationship patterns and characteristics made a company or person high-risk. With the model, the team was able to identify other entities with similar characteristics that had previously not been identified. This has led to new approaches to identifying internal threats and analyzing adversaries.

Read the Blog: Entity Graphs with Novetta Entity Analytics and Amazon Neptune

Audio Analysis Built on Amazon Sagemaker and AWS EC2

While many services exist to help make sense of audio data, they often require that the language be explicitly identified before data is passed to the service. This manual step slows down or even halts the analysis of large amounts of audio data. Novetta overcame this challenge by training deep learning models to automatically identify which language is being spoken in audio files.

Our researchers took a unique approach to this problem, converting the audio signals to images as a new way to extract data, as well as a way to allow for the use of the robust image processing algorithms that are well established in the deep learning community. By taking a creative approach to the problem our researchers developed models that could identify the language being spoken with high accuracy.

Given the large number of audio files that needed to be processed the team used the large, on-demand compute power available on SageMaker and AWS EC2 instances. This enabled them to train resource-intensive Residual Networks and Generative Adversarial Networks (GANs) to perform complex tasks on audio signals. Reducing noise in audio signals required the large compute power of AWS P3 instances to train the networks. These instances make use of 8 NVIDIA V100 Tensor Core GPUs and up to 100 Gbps of networking throughput to handle GAN training data. Access to such powerful machines enabled rapid evaluation of these models.

Read the Blog: Deep Learning for Language Identification on Audio Signals