June '19 Newsletter

AI-Native Software Infrastructure

At Determined AI, we remain hard at work building specialized software infrastructure for deep learning (DL). Since our public launch in March, our team has grown significantly to expedite product development. Below is information on some of our newest features, as well as some recommended reading from our DL experts.

Usability enhancements for experimentation

Several new features have been added to Determined AI to help enhance workflows related to early-stage model development and experimentation. Specifically, Determined AI now offers:

  • Integrated Jupyter Notebooks: Deploy a GPU-attached Jupyter notebook in a single command and easily use the notebook interface to develop and/or debug model definition code, visualize results of trained model(s) or even manage the entire training lifecycle of a model.

Jupyter Notebooks in Action

  • Seamless Tensorboard integration: Regardless of the framework you’re using (TensorBoard, PyTorch, Keras, etc.), you can visualize and compare experiment metrics and to diagnose everything from vanishing gradients to exactly when your models start overfitting.

  • Scheduling arbitrary GPU-backed containers with Commands: Commands allow for increased flexibility in sharing GPU resources while enhancing reproducibility.

  • Simplified configuration management with configuration templates: Reduce redundancy in experiment configuration files by consolidating settings that are shared by many experiments into a single YAML-based template.

We’ll be highlighting several other new product features in the coming weeks, including enhancements to parallel and distributed training, automated deployment optimization, and advanced model serving capabilities.

Add these insightful articles from DL experts to your summer reading list:

  • The cloud giants have an AI problem. Think you can see further by standing on the shoulders of the cloud giants? Think again. Our CEO Evan Sparks explains why cloud computing is not a panacea, especially when it comes to AI modeling, and how it could actually cost you nearly 10x more than on prem alternatives.

  • Stop doing iterative model development. You can spend many cycles on iterative model development to improve a model’s performance over an established baseline. In this blog post, Yoav Zimmerman introduces a new “search-driven model development” paradigm that can shrink months of painstaking, iterative work down to a single overnight job.

  • Random Search is a hard baseline to beat for Neural Architecture Search. Ameet Talwalkar demonstrates the potential promise of Neural Architecture Search (NAS) and the current immaturity of the field in this blog post and presentation at O’Reilly AI Conference.

  • A recipe for training neural networks. Andrej Karpathy advocates for a structured process for neural network training, and provides a step-by-step recipe to prevent “the introduction of a lot of unverified complexity at once.” He offers good tips for every stage of model development lifecycle, with insights that are complementary to those provided in Yoav’s aforementioned post.