The Data Exchange Podcast: Evan Sparks on accelerating the adoption of deep learning in the enterprise.
In this episode of the Data Exchange I speak with Evan Sparks, cofounder and CEO of Determined AI1, a startup that recently open sourced a platform for training deep learning models. Many of the impressive results and applications of deep learning have happened at a handful of companies and research groups. As more companies use deep learning they are learning that infrastructure for training and transfer learning isn’t widely available.
Our conversation focused on deep learning and other topics including:
- Their decision to open source the Determined Training Platform (DTP).
- Enterprise use cases and applications of deep learning, and why Evan thinks more companies will need a platform for training DL models.
- The components that come with the DTP: Distributed Training and Hyperparameter Tuning, Experiment Tracking and tools for collaboration and governance, Scheduler specialized for DL workflows, and more.
- Some examples of how teams have been using DTP.
(Full transcript of our conversation is below.)
Subscribe to our Newsletter:
We also publish a popular newsletter where we share highlights from recent episodes, trends in AI / machine learning / data, and a collection of recommendations.
Download a complete transcript of this episode by filling out the form below:
Ben: Let’s start with the premise behind Determined —what does the company do?
Evan: We build software tools for teams of machine learning engineers who are trying to build deep learning models to help them be much, much more productive, and we offer a software platform that helps them with the training part of their deep learning workflow. It’s kind of like power tools for deep learning model training. We provide tools to accelerate different aspects of that workflow and help data scientists and machine learning engineers ship better models faster than they could before.
Ben: Why should a regular company bother with deep learning when they’re just getting their hands around their data? For example, let’s say a company has a data warehouse with some BI built in, a few dashboards, and now they’re starting to build simple models for forecasting. Why should they prepare for an age where deep learning is critical?
Evan: I think the amazing thing about deep learning is the new types of data in the modalities that it’s allowed us to start modeling, and we think that just about every company has the kinds of super high-value business problems that could be solved with deep learning—where even if they lack the internal sophistication and expertise to build these sorts of models today, they could potentially gain a lot by investing in building that out for targeted use cases.
I’ll give you a few examples. Most of your listeners probably know the areas where deep learning works best—specifically, computer vision and (increasingly) text and NLP-type processing and speech. Again, we think a lot of companies have this sort of data. One example is an insurance company we’ve talked to. They have a massive database of claims data that includes photos of cars that have been involved in wrecks. They want to quickly identify whether or not a car is totaled just by letting a consumer take a picture of their car, then the insurance company can decide how to act on that much more quickly—to know if they need to get a tow truck and have it hauled away, or if they need to send it to the nearest body shop. That sort of thing could not have been imagined 10 years ago by an insurance company. You were literally sending people out into the field to make those sorts of assessments. So the cost savings are amazing.
- Dean Wampler: “Scalable Machine Learning, Scalable Python, For Everyone”
- Rajat Monga: “The evolution of TensorFlow and of machine learning infrastructure”
- Edo Liberty: “How deep learning is being used for search and information retrieval”
- Edmon Begoli: “Hyperscaling natural language processing”
 I am an advisor to Determined AI.
[Image: Edges in Alexander Platz by Dean Wampler, used with permission.]