Determined AI AutoML Helps Manage GPU Clusters

Determined AI, a San Franciso-based startup has built an AutoML platform that provides a comprehensive set of deep learning services including AI workflows, data management, model training, deployment, automation, visualization tools, and GPU cluster management. Co-founders Evan Sparks and Neil Conway earned their Ph.D.’s from Berkeley, and Ameet Talwalkar from New York University, who is also moonlighting as an Assistant Professor at Carnegie Mellon University. Carnegie Mellon is a powerhouse in AI research.

Background

  • Company: Determined AI
  • Founded: 2017
  • Raised: $13.6M
  • HQ: San Francisco
  • # of Employees: 26
  • Founders: Evan Sparks (CEO), Neil Conway (CTO), and Ameet Talwalkar (Chief Scientist)
  • Investors: Google Ventures and others
  • Product: AutoML Platform

The platform supports multiple open source AI frameworks including TensorFlow (Google), Keras, and PyTorch (Facebook). AI workloads can run on bare-metal, shared GPU resources, on-premise, or in the cloud. One key platform feature is the ability to accelerate model development across a distributed training environment. In addition, with just one click, AI developers can access and work with TensorBoard.

Tensorboard, which is a part of the TensorFlow ecosystem is a powerful visualization tool used to experiment with machine learning models. As illustrated below, TensorBoard allows AI developers to view charts like histograms, metrics such as weights, biases, loss, and accuracy, and much more.

source: Tensorboard

One area the startup excels at is hyperparameter optimization, also known as hyperparameter tuning. In layman terms, when an AI engineer is building an ML model,  the parameters in each layer must be built. The graph below shows some of the hyperparameters.

Co-founder Ameet Talwalkar noted at a meetup that tuning hyperparameters could take “days to weeks” due to the need to “evaluate a large number of configurations”. Thus, his team developed an algorithm called ASHA which helps with hyperparameter tuning. Based on extensive empirical data, they concluded that ASHA outperforms Population Based Training and Bayesian Optimization and Hyperband techniques. All in all, Determined AI is the right company at the right time to solve ML biggest problems, while also providing an alternative to Google.

source: TensorBoard