In earlier blog posts we introduced Multimetric Optimization, an advanced feature to optimize a model with more than one metric. It automates the process of finding a frontier of best models, and enables you to make the tradeoff between different metrics that you care about. We are excited today to announce improvements to analyzing and viewing Multimetric experiments with the SigOpt Experiment Insights dashboard.
In this third post in our uncertainty series, we consider uncertainty in the context of multimetric optimization; specifically, we discuss how to interpret the outcome of such problems. Balancing competing metrics requires searching the possible configurations to identify how certain configurations tradeoff high performance in one metric with lower performance in the other. Doing so when observations are uncertain, or noisy, is complicated since the future performance of that configuration can not be perfectly known from the observed performance. This post explores this complexity and suggests strategies for interpreting noisy results to help make reliable decisions regarding which configurations are most preferred.
In our previous blog post, we discussed the importance of properly accounting for uncertainty when constructing machine learning models for regression/classification. Here we focus on how uncertainty plays an important role in Bayesian optimization. We will show how a naive strategy for optimization (random search) falls short in problems with uncertainty in the function values. Then, we will review the fundamentals of Bayesian optimization, highlighting acquisition functions that could be used in this setting. Finally, we illustrate the difference of these approaches, as well as the impact of noisy observations, for a simple optimization problem.
Uncertainty exists and is generally unavoidable. This is neither surprising not intrinsically problematic. But having to make predictions given uncertainty about the world as it has been observed can be difficult. There is a potential to too closely follow observed results and overfit to noise, and there is potential to disbelieve observations in favor of only slightly informed prior beliefs. The ideal models, generally, are somewhere in the middle.
In this blog post, we briefly consider the role of uncertainty in building models. The discussion will focus on the balance between building models that respect the data and models that are well-behaved (according to a pre-defined sense of what that means). In follow-up blog posts we will discuss how these decisions play a role in Bayesian optimization.
Each year, a lot of amazing research is presented at ICML, and specifically the AutoML workshop, by leaders of the automatic machine learning community. Each year that we have attended we have found it to be an exciting exchange of ideas in a friendly community; but this year, held special importance to SigOpt's research team. On July 14, the results of the PAKDD 2018 data competition will be presented and, spoiler alert, not one but TWO alumni of our SigOpt research internship program successfully competed. In this post, we will introduce the competition and give the competitors, Katharina Eggensperger and Jungtaek Kim, a chance to explain their submissions and AutoML strategies.
We are excited to announce the general availability of Constraints, a feature that gives customers more fine-grained control of an experiment's parameter space. Now, a subset of the continuous parameters in an experiment can be subject to a linear constraint to restrict the parameter space that SigOpt searches. It is useful for scenarios where known interdependencies between parameters mean only a region of the parameter space is valid.
Today, we are excited to announce SigOpt Organizations, the next step in the evolution of our web dashboard. As the latest improvement to our web dashboard, Organizations is designed to help larger customers with multiple modelling teams control user access and roll up cross-team usage insights. The result is a more seamless experience for every user, whether it be the chief innovation officer, head of data science, or data scientist.
Today, we’re happy to announce a strategic investment and technology development agreement with In-Q-Tel (IQT).
IQT is a non-profit, strategic investor that helps accelerate the development and delivery of cutting-edge technologies to U.S. government agencies that keep our nation safe. Established in 1999, IQT has the goal of identifying and partnering with startups that are developing innovative technologies that can better protect and preserve the United States’ security. We’re proud to partner with IQT to bring SigOpt’s optimization-as-a-service solution to our nation’s government agencies.
The research team at SigOpt works to provide the best Bayesian optimization platform for our customers. In our spare time, we also engage in research projects in a variety of other fields. This blog post highlights one of those recent projects which will be presented Tuesday February 6 at AAAI 2018. For those who cannot attend that session, we discuss here the topic of embeddings of vectors and the computational gains available when using circulant binary embeddings.
2017 was a great year for SigOpt. We were selected by a number of the leading firms in algorithmic trading, banking, and technology to improve their optimization processes. We announced a number of strategic partnerships with companies like AWS, Intel AI, NVIDIA, and SAP. We won a number of industry awards, such as the Barclays Innovation Challenge Award and the CB Insights AI 100, and were recognized as a Gartner Cool Vendor in AI. We are looking forward to an even better 2018!
Last month, we announced the availability of SigOpt on AWS Marketplace, which allows data scientists and researchers to deploy SigOpt on AWS with the click of a button. Today, we’re doubling down on our collaboration with AWS through a new feature called AWS PrivateLink.
AWS PrivateLink is a new solution that will enable SigOpt to connect directly with any AWS customer that has an Amazon Virtual Private Cloud (VPC). Amazon VPC is a private cloud-based network organizations can leverage to provision an isolated section of the AWS cloud to launch AWS resources.
It's been a busy few months at SigOpt. We've developed and launched new features like Multimetric optimization and Linear Constraints, teamed up with AWS and NVIDIA to publish blog posts, published our research at ICML in August, and more. Read the first edition of our quarterly newsletter to learn about what the team has been working on!
There’s no question that Amazon Web Services (AWS) has played an instrumental role in technology today. AWS has reduced the historically large equipment costs required to build and scale technology—like servers, cables, hard drives, and power supplies—allowing entrepreneurs and software engineers to reap the benefits of the cloud. This is why it is with great pleasure that today we announce the availability of SigOpt on AWS Marketplace.
Here at SigOpt, Gaussian processes and reproducing kernel Hilbert spaces (RKHS) are important components of our Bayesian optimization methodology. Our research into this topic often exists at the intersection of approximation theory and spatial statistics. We recently attended the SIAM Computational Science and Engineering 2017 meeting in Atlanta to continue learning about outstanding accomplishments in these and other exciting topics in computation.
In addition to participating in the career fair, we also organized a minisymposium to bring together experts in Gaussian processes and RKHS and discuss recent advances that are relevant to Bayesian optimization. First, we would like to extend our thanks to Jian Wu, Jie Chen and Greg Fasshauer for contributing talks to the session; their presentations provided a diverse set of insights into new strategies for improving the viability and performance of RKHS theory in Bayesian optimization. This post serves to discuss our ongoing collaboration with Dr. Fasshauer and explore the topic of his talk at a level suitable for a broad audience.
We are releasing a new and improved web experience for analyzing Multimetric experiments!
A discussion on how to interpret Pareto efficiency with noisy observations.
A presentation of different Bayesian optimization strategies for dealing with noisy metrics.
A brief introduction to the impact that uncertainty may have on a robust modeling process.
Discussions regarding a recent ML competition, presented at the AutoML workshop at ICML 2018, in which former SigOpt interns placed first and second.
A geometric interpretation of upper confidence bound presented at ICASSP 2018.
In-depth optimization research