This Amazon intern published a paper that will extend the usability of Amazon SageMaker DeepAR in a profound way

Konstantinos Benidis’ education and professional work experience spans four countries. He has a master’s degree in electrical and computer engineering at the National Technical University of Athens (NTUA), a master’s of science degree from the Universitat Politècnica de Catalunya in Barcelona and the Technical University of Munich, and a PhD from the Hong Kong University of Science and Technology.

While completing his PhD in financial engineering and optimization in Hong Kong, Konstantinos went in-depth into machine learning on his own initiative. At the time, he received internship offers from a variety of companies in the financial and technology sectors. He decided to pursue an opportunity with Amazon.

“I wanted to work at Amazon because I had friends working there. They said that they liked the work environment. They also said they were involved in projects where they could see the end result. This was important to me.”

Sorting out a wrinkle in time series forecasting

Konstantinos joined a team of scientists working on forecasting the distributions of future values of time series. Time series forecasting has applications in numerous scientific fields and commercial applications. For example, the ability to understand complex distributions of data can be used to predict glucose levels in medicine, the future movement of indices in financial markets, and to make accurate forecasts of product demand for inventory management.

These scenarios call for the automation of optimal decision making amidst uncertainty. Producing a point estimate is inadequate for the end goal of making an informed decision. Additionally, traditional forecasting techniques assume a Gaussian distribution—a choice often based on mathematical convenience rather than evidence, rendering the approach inadequate for the real world.

Konstantinos’ team proposed a new approach. SQF-RNN combined the forecasting capacity of recurrent neural networks (RNNs) with the flexibility of a quantile function-based specification of the observation distribution. In a break from the past, the new methodology looked at the spline-based representation of the entire conditional quantile function. This differed from previous approaches that typically focused on modeling only a fixed set of quantiles.

Publishing at Amazon

Prior to joining Amazon, Konstantinos had heard that publishing papers was discouraged within the company. This can be true for confidential projects, however, Konstantinos was surprised to find an atmosphere that encouraged publishing.

“I like structure in my work,” says Konstantinos. “I appreciated having a clear set of goals for my six-month internship. Publishing a paper related to my work was one of the goals that was set out for me.”

The paper published by Konstantinos’ team, Probabilistic Forecasting with Spline Quantile Function RNNs, was accepted at the Artificial Intelligence and Statistics (AISTATS) 2019 conference.

In addition to publishing papers, Konstantinos had been excited at the possibility of seeing his work have real-world impact. His team’s work came to life with the release of the open source Gluon Time Series (GluonTS), a Python toolkit for building, evaluating, and comparing deep learning–based time series models. GluonTS is based on the Gluon interface to Apache MXNet, and provides components that make building time series models simple and efficient.

I appreciated having a clear set of goals for my six-month internship. Publishing a paper related to my work was one of the goals that was set out for me.

Konstantinos completed his PhD even as he worked as an Amazon intern. Following his internship, he decided to move on to a full-time role within the company. As a full-time employee, Konstantinos is actively working on publishing more papers, in addition to gaining management experience. One of his responsibilities today involves supervising an intern.

Konstantinos is excited by the possibility of seeing his work have an impact on multiple fronts.

“Amazon is creating entirely new ways of doing things in so many areas. When you combine this with the fact that you can collaborate with people across multiple teams, it makes coming to work a fulfilling experience every day.”