SigOpt Automated Model Tuning: Product Overview and Analysis

SigOpt's Optimization Solution automates the tuning of any model built with any framework on any infrastructure to maximize the return on machine learning, artificial intelligence and general research investments.

PO_SigOpt

eWEEK is compiling a new IT products and services section that encompasses most of the categories that we cover on our site. In it, we will spotlight the leaders in each sector, which include enterprise software, hardware, security, on-premises-based systems and cloud services. We also will add promising new companies as they come into the market.

Company:  SigOpt (automated data model tuning for artificial intelligence and machine-learning platforms)

Company description: SigOpt's optimization solution automates the tuning of any model built with any framework on any infrastructure to maximize the return on machine learning, artificial intelligence and general research investments. Built by experts for experts, this solution embeds an ensemble of Bayesian and global optimization algorithms within a standardized platform that is accessible through a simple REST API. This automated, scalable and comprehensive approach enables teams to tune much earlier and more often, which, in turn, helps transform traditional discrete data projects with mathematical outputs to continuously deployed products that improve business outcomes.

Founded in 2014, SigOpt was born out of the desire to make experts more efficient. While co-founder and CEO Scott Clark was completing his Ph.D at Cornell, he noticed that often the final stage of research was a domain expert tweaking what they had built via trial and error. After completing his doctorate, Clark developed MOE to solve this problem and used it to optimize machine learning models and A/B tests (sometimes called split testing) at Yelp.

SigOpt is a privately held company and is supported by leading investors from Andreessen Horowitz, DCVC, Y Combinator, SV Angel, Blumberg Capital, and In-Q-Tel, and has won awards from O'Reilly, Barclay's, CB Insights, and Gartner. Its academic community includes users from MIT, Stanford University, Cal-Berkeley, Cornell, Carnegie Mellon, Harvard and other leading institutions. Partners include Amazon, Google, Intel, NVIDIA and other leaders in AI.

SigOpt is located at 100 Bush St, San Francisco, Calif. 94104.

Markets: Global

International Operations: SigOpt does not have offices overseas.

Product and Services: The SigOpt Optimization Platform is a B2B cloud SaaS solution that automates the tuning of any data model built with any framework on any infrastructure to maximize the return on machine learning, artificial intelligence and general research investments. Built by experts for experts, this solution embeds an ensemble of Bayesian and global optimization algorithms within a standardized platform that is accessible through a simple REST application programming interface (API).

Key Features: SigOpt’s Optimization Solution has three capabilities that empower teams to tune any data model cheaper, better and faster:

  • Experiment Insights: Uncover cross-experiment trends or introspect particular experiments to make data-driven decisions that improve the model development process
  • Optimization Engine: Automate hyperparameter optimization for any model with an ensemble of algorithms capable of tuning any volume, variety, or complexity of models
  • Enterprise-Grade Platform: Implement the solution with only 20 lines of code and be confident that it will reliably scale with your machine-learning needs

Teams take advantage of these capabilities with a simple process that is all enabled with SigOpt’s REST API:

  • You provide SigOpt with your parameters: You ping the API with your model’s parameters; SigOpt doesn’t need the model itself. You keep it private.
  • You use SigOpt's values: The API suggests new values for these parameters. Use them to evaluate your model within your current infrastructure.
  • You send your model output: SigOpt uses your model’s output to calculate the next best configuration.
  • Repeat until optimized: You’ll reach optimal values up to 100x faster than other methods.

SigOpt’s black-box optimization allows the company to tune your models without accessing them or the underlying data. This lightweight approach means that you control the processes--data and modeling--that directly benefit from data scientist expertise, while outsourcing to SigOpt the optimization process that does not benefit from expertise. This allocation of expert time ensures more productive and effective machine learning that delivers significant return on any team’s machine learning investment.

Insight and Analysis:  There is little, if any, third-party analysis available on SigOpt’s software. Gartner Peer Insights and IT Central Station have no reviews as of Aug. 3, 2018.

There is a deep, valuable look at SigOpt research, methodology and parameter optimization available here. The company lays out its mission and discussion of projects and the market in-in-depth fashion in its blog here.

SigOpt is available in the AWS Marketplace.

Nvidia Developer has a good explanation of how to use SigOpt’s Bayesian optimization platform here.

List of current customers: More than a dozen enterprise customers as well as hundreds of academics currently deploy SigOpt. Customers include Quantopian, In-Q-Tel, Hotwire, MIT and WorkFusion.

Delivery: The SigOpt Optimization Platform is a B2B cloud software as a service solution, whose API is easy to stand up with just a few lines of code and requires no implementation partners. It works with all of public clouds and offers private on-premises solutions for any customers who require this.

Pricing:  SigOpt’s Optimization Solution can be accessed following a 30-day free trial, or as monthly licenses billed annually (the cost of which is determined by multiple factors including number of seats, number of experiments per month, number of observations per experiment, and number of parameters per experiment). SigOpt is free to academics.

AWS carries the following pricing information:

Units

Description

1 month

1 year

3 years

15 experiments/month

With 500 observations, 20 parameters, 4x parallelism

$3,000

$30,000

$75,000

100 experiments/month

With 1000 observations, 100 parameters, 10x parallelism

$60,000

$120,000

$300,000

For more information about pricing, go here.

Other key players in this market: Google CloudML, Google AutoML, AWS SageMaker

Contact information for potential customers:  Contact us or find out more about SigOpt here.

Resources:

SigOpt Research backgrounder

AWS listing and backgrounder

Crunchbase

If you have a suggestion for an eWEEK Product Overview and Analysis article, email cpreimesberger@eweek.com.

Chris Preimesberger

Chris J. Preimesberger

Chris J. Preimesberger is Editor-in-Chief of eWEEK and responsible for all the publication's coverage. In his 13 years and more than 4,000 articles at eWEEK, he has distinguished himself in reporting...