Social
Vancouver, Canada
Machine learning

3 important themes from Amazon's 2019 NeurIPS papers

Time series forecasting, bandit problems, and optimization are integral to Amazon's efforts to deliver better value for its customers.

By Larry Hardesty
November 25, 2019
14 min read
Share

Last year, the first 2,000-2,500 publicly released tickets to the Conference on Neural Information Processing Systems, or NeurIPS, sold out in 12 minutes.

This year, the conference organizers moved to a lottery system, allowing aspiring attendees to register in advance and randomly selecting invitees from the pool of registrants. But they also bumped the number of public-release tickets up from around 2,000 to 3,500, testifying to the conference’s continued popularity.

At NeurIPS this year, there are 26 papers with Amazon coauthors. They cover a wide range of topics, but surveying their titles, Alex Smola, a vice president and distinguished scientist in the Amazon Web Services organization, discerns three prominent themes, all tied to Amazon’s efforts to deliver better value for its customers.

Those three themes are time series forecasting (and causality), bandit problems, and optimization.

1. Time series forecasting

Time series forecasting involves measuring some quantity over time — such as the number of deliveries in a particular region in the past six months, or the number of cloud servers required to support a particular site over the past two years — and attempting to project that quantity into the future.

“That’s something that is very dear to Amazon’s heart,” Smola says. “For anything that Amazon does, it’s really beneficial to have a good estimate of what our customers will expect from us ahead of time. Only by being able to do that will we be able to satisfy customers’ demands, be it for products or services.”

A sequence of basis time series, forecast into the near future and summed together to approximate a new time series.
The paper “Think Globally, Act Locally” examines data sets with many correlated time series, such as the demand curves for millions of products sold online. The researchers describe a method for constructing a much smaller set of “basis time series”; the time series for any given product can be approximated by a weighted sum of the bases.
Courtesy of the researchers

The basic mathematical framework for time series forecasting is a century old, but the scale of modern forecasting problems calls for new analytic techniques, Smola says.

“Problems are nowadays highly multivariate,” Smola says. “If you look at the many millions of products that we offer, you want to be able to predict fairly well what will sell, where and to whom.

“You need to make reasonable assumptions on how this very large problem can be decomposed into smaller, more tractable pieces. You make structural approximations, and sometimes those structural approximations are what leads to very different algorithms.

“So you might, for instance, have a global model, and then you have local models that address the specific items or address the specific sales. If you look at ‘Think Globally, Act Locally’” — a NeurIPS paper whose first author is Rajat Sen, an applied scientist in the Amazon Search group — “it’s already in the title. Or look at ‘High-Dimensional Multivariate Forecasting with Low-Rank Gaussian Copula Processes’. In this case, you have a global structure, but it’s only in a small subspace where interesting things happen.”

Side-by-side images depict correlations between taxi traffic at different points in Manhattan at different times of day
The paper "High-Dimensional Multivariate Forecasting with Low-Rank Gaussian Copula Processes" describes a method for predicting correlations among many parallel time series. In one example, the researchers forecast correlations between the taxi traffic at different points in New York City at different times of day. Red lines indicate strong correlations; blue lines indicate strong negative correlations. Weekend midday traffic patterns (left) show negative correlations between locations near the Empire State Building, suggesting that taxis tend to prefer different routes depending on traffic conditions. Weekend evening traffic patterns show positive correlations between the vicinity of the Empire State Building and areas with high concentrations of hotels.
Courtesy of the researchers

An aspect of forecasting that has recently been drawing more attention, Smola says, is causality. Where traditional machine learning models merely infer statistical correlations between data points, “it is ultimately the causal relationship that matters,” Smola says.

“I think that causality is one of the most interesting conceptual developments affecting modern machine learning,” says Bernhard Schölkopf, like Smola a vice president and distinguished scientist in Amazon Web Services. “This is the main topic that I have been interested in for the last decade.”

Two of Schölkopf’s NeurIPS papers — “Perceiving the Arrow of Time in Autoregressive Motion” and “Selecting Causal Brain Features with a Single Conditional Independence Test per Feature” — address questions of causality, as does “Causal Regularization”, a paper by Dominik Janzing, a senior research scientist in Smola’s group.

“Normal machine learning builds on correlations of other statistical dependences,” Schölkopf explains. “This is fine as long as the source of the data doesn't change. For example, if in the training set of an image recognition system, all cows are standing on green pasture, then it is fine for an ML system to use the green as a useful feature in recognizing cows, as long as the test set looks the same. If in the test set, the cows are standing on the beach, then such a purely statistical system can fail.

“More generally: causal learning and inference attempts to understand how systems respond to interventions and other changes, and not just how to predict data that looks more or less the same as the training data.”

2. Bandit problems

The second major theme that Smola discerns in Amazon scientists’ NeurIPS papers is a concern with bandit problems, a phrase that shows up in the titles of Amazon papers such as “MaxGap Bandit: Adaptive Algorithms for Approximate Ranking” and “Low-Rank Bandit Methods for High-Dimensional Dynamic Pricing”. Bandit problems take their name from one-armed bandits, or slot machines.

“It used to be that those bandits were all mechanical, so there would be slight variations between them, and some would have maybe a slightly a higher return than others,” Smola explains. “I walk into a den of iniquity, and I want to find the one-armed bandit where I will lose the least money or maybe make some money. And the only feedback I have is that I pull arms, and I get money or lose money. These are very unreliable, noisy events.”

Bandit problems present what’s known as an explore-exploit trade-off. The gambler must simultaneously explore the environment — determine which machines pay out the most — and exploit the resulting knowledge — concentrate as much money as possible on the high-return machines. Early work on bandit problems concerned identifying the high-return machines with minimal outlays.

“That problem was solved about 20 years ago,” Smola says. “What hasn’t been solved — and this is where things get a lot more interesting — is once you start adding context. Imagine that I get to show you various results as you’re searching for your next ugly Christmas sweater. The unfortunate thing is that the creativity of sweater designers is larger than what you can fit on a page. Now the context is essentially, what time, where from, which user, all those things. We want to find and recommend the ugly Christmas sweater that works specifically for you. This is an example where context is immediately relevant.”

It’s really beneficial to have a good estimate of what our customers will expect from us ahead of time. Only by being able to do that will we be able to satisfy customers’ demands.
Alex Smola, VP and distinguished scientist, Amazon

In the bandit-problem framework, in other words, the high-payout machines change with every new interaction. But there may be external signals that indicate how they’re changing.

Distributed computing, which is inescapable for today’s large websites, changes the structure of the bandit problem, too.

“Say you go to a restaurant, and the cook wants to improve the menu,” Smola says. “You can try out lots of new menu items, and that’s a good way to improve the menu overall. But if you start offering a lot of undercooked dishes because you’re experimenting, then at some point your loyal customers will stay away.

“Now imagine you have 100 restaurants, and they all do the same thing at the same time. They can’t necessarily communicate at the per-second level; maybe every day or every week they chat with each other. Now this entire exploration problem becomes a little more challenging, because if two restaurants try out the same undercooked dish, you make the customer less happy than you could have.

“So how does this map back into Amazon land? Well, if you have many servers doing this recommendation, the explore-exploit trade-off might be too aggressive if every one of them works on their own.”

3. Optimization

Finally, Smola says, “There is a third category of results that has to do with making algorithms faster. If you look at ‘Primal-Dual Block Frank-Wolfe’, ‘Communication-Efficient Distributed SGD with Sketching’, ‘Qsparse-Local-SGD’ — those are the workhorses that run underneath all of this. Making them more efficient is obviously something that we care about, so we can respond to customer requests faster, train algorithms faster.”

Bird’s-eye view

NeurIPS is a huge conference, with more than 1,400 accepted papers that cover a bewildering variety of topics. Beyond the Amazon papers, Caltech professor and Amazon fellow Pietro Perona identifies three research areas as growing in popularity.

“One is understanding how deep networks work, so that we can better design architectures and optimization algorithms to train models,” Perona says. “Another is low-shot learning. Machines are still much less efficient than humans at learning, in that they need more training examples to achieve the same performance. And finally, AI and society — identifying opportunities for social good, sustainable development, and the like.”

NeurIPS is being held this year at the Vancouver Convention Center, and the main conference runs from Dec. 8 to Dec. 12. The Women in Machine Learning Workshop, for which Amazon is a gold-level sponsor, takes place on Dec. 9; the Third Conversational AI workshop, whose organizers include Alexa AI principal scientist Dilek Hakkani-Tür, will be held on Dec. 14.

Amazon's involvement at NeurIPS

Paper and presentation schedule

Tuesday, 12/10 | 10:45-12:45pm | East Exhibition Hall B&C

A Meta-MDP Approach to Exploration for Lifelong Reinforcement Learning | #192
Francisco Garcia (UMass Amherst/Amazon) · Philip Thomas (UMass Amherst)

Blocking Bandits | #17
Soumya Basu (UT Austin) · Rajat Sen (UT Austin/Amazon) · Sujay Sanghavi (UT Austin/Amazon) · Sanjay Shakkottai (UT Austin)

Causal Regularization | #180
Dominik Janzing (Amazon)

Communication-Efficient Distributed SGD with Sketching | #81
Nikita Ivkin (Amazon) · Daniel Rothchild (University of California, Berkeley) · Md Enayat Ullah (Johns Hopkins University) · Vladimir Braverman (Johns Hopkins University) · Ion Stoica (UC Berkeley) · Raman Arora (Johns Hopkins University)

Learning Distributions Generated by One-Layer ReLU Networks | #49
Shanshan Wu (UT Austin) ·Alexandros G. Dimakis (UT Austin) · Sujay Sanghavi (UT Austin/Amazon)

Tuesday, 12/10 | 5:30-7:30pm | East Exhibition Hall B&C

Efficient Communication in Multi-Agent Reinforcement Learning via Variance Based Control | #195
Sai Qian Zhang (Harvard University) · Qi Zhang (Amazon) · Jieyu Lin (University of Toronto)

Extreme Classification in Log Memory using Count-Min Sketch: A Case Study of Amazon Search with 50M Products | #37
Tharun Kumar Reddy Medini (Rice University) · Qixuan Huang (Rice University) · Yiqiu Wang (Massachusetts Institute of Technology) · Vijai Mohan (Amazon) · Anshumali Shrivastava (Rice University/Amazon)

Iterative Least Trimmed Squares for Mixed Linear Regression | #50
Yanyao Shen (UT Austin) · Sujay Sanghavi (UT Austin/Amazon)

Meta-Surrogate Benchmarking for Hyperparameter Optimization | #6
Aaron Klein (Amazon) · Zhenwen Dai (Spotify) · Frank Hutter (University of Freiburg) · Neil Lawrence (University of Cambridge) · Javier Gonzalez (Amazon)

Qsparse-local-SGD: Distributed SGD with Quantization, Sparsification and Local Computations | #32
Debraj Basu (Adobe) · Deepesh Data (UCLA) · Can Karakus (Amazon) · Suhas Diggavi (UCLA)

Selecting Causal Brain Features with a Single Conditional Independence Test per Feature | #139
Atalanti Mastakouri (Max Planck Institute for Intelligent Systems) · Bernhard Schölkopf (MPI for Intelligent Systems/Amazon) · Dominik Janzing (Amazon)

Wednesday, 12/11 | 10:45-12:45pm | East Exhibition Hall B&C

On Single Source Robustness in Deep Fusion Models | #93
Taewan Kim (Amazon) · Joydeep Ghosh (UT Austin)

Perceiving the Arrow of Time in Autoregressive Motion | #155
Kristof Meding (University Tübingen) · Dominik Janzing (Amazon) · Bernhard Schölkopf (MPI for Intelligent Systems/Amazon) · Felix A. Wichmann (University of Tübingen)

Wednesday, 12/11 | 5:00-7:00pm | East Exhibition Hall B&C

Compositional De-Attention Networks | #127
Yi Tay (Nanyang Technological University) · Anh Tuan Luu (MIT) · Aston Zhang (Amazon) · Shuohang Wang (Singapore Management University) · Siu Cheung Hui (Nanyang Technological University)

Low-Rank Bandit Methods for High-Dimensional Dynamic Pricing | #3
Jonas Mueller (Amazon) · Vasilis Syrgkanis (Microsoft Research) · Matt Taddy (Amazon)

MaxGap Bandit: Adaptive Algorithms for Approximate Ranking | #4
Sumeet Katariya (Amazon/University of Wisconsin-Madison) · Ardhendu Tripathy (UW Madison) · Robert Nowak (UW Madison)

Primal-Dual Block Generalized Frank-Wolfe | #165
Qi Lei (UT Austin) · Jiacheng Zhuo (UT Austin) · Constantine Caramanis (UT Austin) · Inderjit S Dhillon (Amazon/UT Austin) · Alexandros Dimakis (UT Austin)

Towards Optimal Off-Policy Evaluation for Reinforcement Learning with Marginalized Importance Sampling | #208
Tengyang Xie (University of Illinois at Urbana-Champaign) · Yifei Ma (Amazon) · Yu-Xiang Wang (UC Santa Barbara)

Thursday, 12/12 | 10:45-12:45pm | East Exhibition Hall B&C

AutoAssist: A Framework to Accelerate Training of Deep Neural Networks | #155
Jiong Zhang (UT Austin) · Hsiang-Fu Yu (Amazon) · Inderjit S Dhillon (UT Austin/Amazon)

Exponentially Convergent Stochastic k-PCA without Variance Reduction | #200 (oral, 10:05-10:20 W Ballroom C)
Cheng Tang (Amazon)

Failing Loudly: An Empirical Study of Methods for Detecting Dataset Shift | #54
Stephan Rabanser (Technical University of Munich/Amazon) · Stephan Günnemann (Technical University of Munich) · Zachary Lipton (Carnegie Mellon University/Amazon)

High-Dimensional Multivariate Forecasting with Low-Rank Gaussian Copula Processes | #107
David Salinas (Naverlabs) · Michael Bohlke-Schneider (Amazon) · Laurent Callot (Amazon) · Jan Gasthaus (Amazon) · Roberto Medico (Ghent University)

Learning Search Spaces for Bayesian Optimization: Another View of Hyperparameter Transfer Learning | #30
Valerio Perrone (Amazon) · Huibin Shen (Amazon) · Matthias Seeger (Amazon) · Cedric Archambeau (Amazon) · Rodolphe Jenatton (Amazon)

Mo’States Mo’Problems: Emergency Stop Mechanisms from Observation | #227
Samuel Ainsworth (University of Washington) · Matt Barnes (University of Washington) · Siddhartha Srinivasa (University of Washington/Amazon)

Think Globally, Act Locally: A Deep Neural Network Approach to High-Dimensional Time Series Forecasting | #113
Rajat Sen (Amazon) · Hsiang-Fu Yu (Amazon) · Inderjit S Dhillon (UT Austin/Amazon)

Thursday, 12/12 | 5:00-7:00pm | East Exhibition Hall B&C

Dynamic Local Regret for Non-Convex Online Forecasting | #20
Sergul Aydore (Stevens Institute of Technology) · Tianhao Zhu (Stevens Institute of Technology) · Dean Foster (Amazon)

Interaction Hard Thresholding: Consistent Sparse Quadratic Regression in Sub-quadratic Time and Space | #47
Suo Yang (UT Austin), Yanyao Shen (UT Austin), Sujay Sanghavi (UT Austin/Amazon)

Inverting Deep Generative Models, One Layer at a Time |#48
Qi Lei (University of Texas at Austin) · Ajil Jalal (UT Austin) · Inderjit S Dhillon (UT Austin/Amazon) · Alexandros Dimakis (UT Austin)

Provable Non-linear Inductive Matrix Completion| #215
Kai Zhong (Amazon) · Zhao Song (UT Austin) · Prateek Jain (Microsoft Research) · Inderjit S Dhillon (UT Austin/Amazon)

Amazon researchers on NeurIPS committees and boards

  • Bernhard Schölkopf – Advisory Board
  • Michael I. Jordan – Advisory Board
  • Thorsten Joachims – senior area chair
  • Anshumali Shrivastava – area chair
  • Cedric Archambeau – area chair
  • Peter Gehler – area chair
  • Sujay Sanghavi – committee member

Workshops

Learning with Rich Experience: Integration of Learning Paradigms

Paper: "Meta-Q-Learning" | Rasool Fakoor, Pratik Chaudhari, Stefano Soatto, Alexander J. Smola

Human-Centric Machine Learning

Paper: "Learning Fair and Transferable Representations" | Luco Oneto, Michele Donini, Andreas Maurer, Massimiliano Pontil

Bayesian Deep Learning

Paper: "Online Bayesian Learning for E-Commerce Query Reformulation" | Gaurush Hiranandani, Sumeet Katariya, Nikhil Rao, Karthik Subbian

Meta-Learning

Paper: "Constrained Bayesian Optimization with Max-Value Entropy Search" | Valerio Perrone, Iaroslav Shcherbatyi, Rodolphe Jenatton, Cedric Archambeau, Matthias Seeger

Paper: "A Quantile-Based Approach to Hyperparameter Transfer Learning" | David Salinas, Huibin Shen, Valerio Perrone

Paper: "A Baseline for Few-Shot Image Classification" | Guneet Singh Dhillon, Pratik Chaudhari, Avinash Ravichandran, Stefano Soatto

Conversational AI

Organizer: Dilek Hakkani-Tür

Paper: "The Eighth Dialog System Technology Challenge" | Seokhwan Kim, Michel Galley, Chulaka Gunasekara, Sungjin Lee, Adam Atkinson, Baolin Peng, Hannes Schulz, Jianfeng Gao, Jinchao Li, Mahmoud Adada, Minlie Huang, Luis Lastras, Jonathan K. Kummerfeld, Walter S. Lasecki, Chiori Hori, Anoop Cherian, Tim K. Marks, Abhinav Rastogi, Xiaoxue Zang, Srinivas Sunkara, Raghav Gupta

Paper: “Just Ask: An Interactive Learning Framework for Vision and Language Navigation” | Ta-Chung Chi, Minmin Shen, Mihail Eric, Seokhwan Kim, Dilek Hakkani-Tur

Paper: “MA-DST: Multi-Attention-Based Scalable Dialog State Tracking” | Adarsh Kumar, Peter Ku, Anuj Kumar Goyal, Angeliki Metallinou, Dilek Hakkani-Tür

Paper: “Investigation of Error Simulation Techniques for Learning Dialog Policies for Conversational Error Recovery” | Maryam Fazel-Zarandi, Longshaokan Wang, Aditya Tiwari, Spyros Matsoukas

Paper: “Towards Personalized Dialog Policies for Conversational Skill Discovery”| Maryam Fazel-Zarandi, Sampat Biswas, Ryan Summers, Ahmed Elmalt, Andy McCraw, Michael McPhillips, John Peach

Paper: “Conversation Quality Evaluation via User Satisfaction Estimation” | Praveen Kumar Bodigutla, Spyros Matsoukas, Lazaros Polymenakos

Paper: “Multi-domain Dialogue State Tracking as Dynamic Knowledge Graph Enhanced Question Answering” | Li Zhou, Kevin Small

Science Meets Engineering of Deep Learning

Paper: "X-BERT: eXtreme Multi-label Text Classification using Bidirectional Encoder from Transformers" Wei-Cheng Chang, Hsiang-Fu Yu, Kai Zhong, Yiming Yang, Inderjit S. Dhillon

Machine Learning with Guarantees

Organizers: Ben London, Thorsten Joachims
Program Committee: Kevin Small, Shiva Kasiviswanathan, Ted Sandler

MLSys: Workshop on Systems for ML

Paper: "Block-Distributed Gradient Boosted Trees" | Theodore Vasiloudis, Hyunsu Cho, Henrik Boström

Women in Machine Learning

Gold sponsor: Amazon

Research areas
Tags
About the Author
Larry Hardesty
Larry Hardesty is the editor of the Amazon Science blog. Previously, he was a senior editor at MIT Technology Review and the computer science writer at the MIT News Office.

Related content

Work with us

See more jobs See more jobs
Data Scientist II, Last Mile Science
IN, KA, Bangalore
Does the thought of improving one of the world’s most complex logistic systems inspire you? Is your passion to sift through hundreds of systems, processes, and data sources to solve the puzzle and identify the next big opportunity? Are you a creative big thinker who is passionate about using data to direct decision making and solve complex and large-scale challenges? Are you fascinated by the interactions between operations and strategy? Do you feel like your skills uniquely qualify you to bridge communication between teams with competing priorities? If so, then this position is for you! Come help Amazon create state-of-the-art science-driven technologies for delivering packages to the doorstep of our customers! The Last Mile Routing & Planning organization builds the software, algorithms and tools that make the “magic” of home delivery happen: our flow, sort, dispatch and routing intelligence systems are responsible for the billions of daily decisions needed to plan and execute safe, efficient and frustration-free routes for drivers around the world. Our team supports deliveries (and pickups!) for Amazon Logistics, Same Day, Amazon Grocery, Lockers, and other new initiatives across the world. Key job responsibilities In this role, your main focus will be to apply algorithms, synthesize information, identify business opportunities, provide data-driven insights and communicate business and technical requirements within the team and across stakeholder groups. You will partner closely with other scientists and engineers in a collegial environment with a clear path to business impact. We have an exciting portfolio of research areas including vehicle routing, planning for electric and autonomous vehicles, district and stops planning, ultra-fast deliveries, fleet planning, and forecasting solutions for different delivery programs leveraging the latest OR, ML, and Generative AI methods, at a global scale. Successful candidates will have a deep knowledge of Operations Research and/or Machine/Deep Learning methods, experience in applying these methods to large-scale business problems, the ability to map models into production-worthy code in Python or Java, the communication skills necessary to explain complex technical approaches to a variety of stakeholders and customers, and the excitement to take iterative approaches to tackle big research challenges.
Applied Scientist, Fauna
US, NY, New York
We are seeking an Applied Scientist to lead the development of evaluation frameworks and data collection protocols for robotic capabilities. In this role, you will focus on designing how we measure, stress-test, and improve robot behavior across a wide range of real-world tasks. Your work will play a critical role in shaping how policies are validated and how high-quality datasets are generated to accelerate system performance. You will operate at the intersection of robotics, machine learning, and human-in-the-loop systems, building the infrastructure and methodologies that connect teleoperation, evaluation, and learning. This includes developing evaluation policies, defining task structures, and contributing to operator-facing interfaces that enable scalable and reliable data collection. The ideal candidate is highly experimental, systems-oriented, and comfortable working across software, robotics, and data pipelines, with a strong focus on turning ambiguous capability goals into measurable and actionable evaluation systems. Key job responsibilities - Design and implement evaluation frameworks to measure robot capabilities across structured tasks, edge cases, and real-world scenarios - Develop task definitions, success criteria, and benchmarking methodologies that enable consistent and reproducible evaluation of policies - Create and refine data collection protocols that generate high-quality, task-relevant datasets aligned with model development needs - Build and iterate on teleoperation workflows and operator interfaces to support efficient, reliable, and scalable data collection - Analyze evaluation results and collected data to identify performance gaps, failure modes, and opportunities for targeted data collection - Collaborate with engineering teams to integrate evaluation tooling, logging systems, and data pipelines into the broader robotics stack - Stay current with advances in robotics, evaluation methodologies, and human-in-the-loop learning to continuously improve internal approaches - Lead technical projects from conception through production deployment - Mentor junior scientists and engineers About the team Fauna Robotics, an Amazon company, is building capable, safe, and genuinely delightful robots for everyday life. Our goal is simple: make robots people actually want to live and interact with in everyday human spaces. We believe that future won’t arrive until building for robotics becomes far more accessible. Today, too much effort is spent reinventing the fundamentals. We’re changing that by developing tightly integrated hardware and software systems that make it faster, safer, and more intuitive to create real-world robotic products. Our work spans the full stack: mechanical design, control systems, dynamic modeling, and intelligent software. The focus is not just functionality, but experience. We’re building robots that feel responsive, expressive, and genuinely useful. At Fauna, you’ll work at the frontier of this space, helping define how robots move, manipulate, and interact with people in natural environments. It’s an opportunity to solve hard problems across hardware and software with a team focused on making robotics accessible and joyful to build. If you care about making robotics real for everyone and building systems that are as delightful as they are capable, we’re interested in hearing from you.
Applied Scientist II, Foundation Model
US, CA, Sunnyvale
Industrial is seeking exceptional talent to help develop the next generation of advanced robotics systems that will transform automation at Amazon's scale. We're building revolutionary robotic systems that combine innovative AI, sophisticated control systems, and advanced mechanical design to create adaptable automation solutions capable of working safely alongside humans in dynamic environments. This is a unique opportunity to shape the future of robotics and automation at unprecedented scale, working with world-class teams pushing the boundaries of what's possible in robotic manipulation, locomotion, and human-robot interaction. This role presents an opportunity to shape the future of robotics through innovative applications of deep learning and large language models. We leverage advanced robotics, machine learning, and artificial intelligence to solve complex operational challenges at unprecedented scale. Our fleet of robots operates across hundreds of facilities worldwide, working in sophisticated coordination to fulfill our mission of customer excellence. We are pioneering the development of robotics foundation models that: - Enable unprecedented generalization across diverse tasks - Integrate multi-modal learning capabilities (visual, tactile, linguistic) - Accelerate skill acquisition through demonstration learning - Enhance robotic perception and environmental understanding - Streamline development processes through reusable capabilities The ideal candidate will contribute to research that bridges the gap between theoretical advancement and practical implementation in robotics. You will be part of a team that's revolutionizing how robots learn, adapt, and interact with their environment. Join us in building the next generation of intelligent robotics systems that will transform the future of automation and human-robot collaboration. As an Applied Scientist, you will develop and improve machine learning systems that help robots perceive, reason, and act in real-world environments. You will leverage state-of-the-art models (open source and internal research), evaluate them on representative tasks, and adapt/optimize them to meet robustness, safety, and performance needs. You will invent new algorithms where gaps exist. You’ll collaborate closely with research, controls, hardware, and product-facing teams, and your outputs will be used by downstream teams to further customize and deploy on specific robot embodiments. Key job responsibilities As an Applied Scientist in the Foundations Model team, you will: - Leverage state-of-the-art models for targeted tasks, environments, and robot embodiments through fine-tuning and optimization. - Execute rapid, rigorous experimentation with reproducible results and solid engineering practices, closing the gap between sim and real environments. - Build and run capability evaluations/benchmarks to clearly profile performance, generalization, and failure modes. - Contribute to the data and training workflow: collection/curation, dataset quality/provenance, and repeatable training recipes. - Write clean, maintainable, well commented and documented code, contribute to training infrastructure, create tools for model evaluation and testing, and implement necessary APIs - Stay current with latest developments in foundation models and robotics, assist in literature reviews and research documentation, prepare technical reports and presentations, and contribute to research discussions and brainstorming sessions. - Work closely with senior scientists, engineers, and leaders across multiple teams, participate in knowledge sharing, support integration efforts with robotics hardware teams, and help document best practices and methodologies.
Manager, Research Science, Special Projects
US, CA, San Francisco
Innovators wanted! Are you an entrepreneur? A builder? A dreamer? This role is part of an Amazon Special Projects team that takes the company’s Think Big leadership principle to the extreme. We focus on creating entirely new products and services with a goal of positively impacting the lives of our customers. No industries or subject areas are out of bounds. If you’re interested in innovating at scale to address big challenges in the world, this is the team for you. Here at Amazon, we embrace our differences. We are committed to furthering our culture of inclusion. We have thirteen employee-led affinity groups, reaching 40,000 employees in over 190 chapters globally. We are constantly learning through programs that are local, regional, and global. Amazon’s culture of inclusion is reinforced within our 16 Leadership Principles, which remind team members to seek diverse perspectives, learn and be curious, and earn trust. Our team highly values work-life balance, mentorship and career growth. We believe striking the right balance between your personal and professional life is critical to life-long happiness and fulfillment. We care about your career growth and strive to assign projects and offer training that will challenge you to become your best.
Applied Scientist, Sponsored Products and Brands
US, WA, Seattle
About Sponsored Products and Brands The Sponsored Products and Brands (SPB) team at Amazon Ads is re-imagining the advertising landscape through generative AI technologies, revolutionizing how millions of customers discover products and engage with brands across Amazon.com and beyond. We are at the forefront of re-inventing advertising experiences, bridging human creativity with artificial intelligence to transform every aspect of the advertising lifecycle from ad creation and optimization to performance analysis and customer insights. We are a passionate group of innovators dedicated to developing responsible and intelligent AI technologies that balance the needs of advertisers, enhance the shopping experience, and strengthen the marketplace. If you're energized by solving complex challenges and pushing the boundaries of what's possible with AI, join us in shaping the future of advertising. About our team The General Shopping Intelligence (GSI) team is a highly motivated, collaborative, and fun-loving group with a strong entrepreneurial spirit and bias for action. We provide advanced real-time machine learning services that connect shoppers with the right ads across all platforms and surfaces worldwide. Through deep understanding of both shoppers and products, we help shoppers discover new products they love, enable advertisers to reach their customers most efficiently, and help Amazon continuously innovate on behalf of all customers. We are seeking a motivated Applied Scientist who loves to innovate at the intersection of customer experience, deep learning, generative AI and high-scale machine learning systems. If you're energized by solving complex challenges and pushing the boundaries of what's possible with AI, join us in shaping the future of advertising. Key job responsibilities As an Applied Scientist, you will: * Leverage Generative AI and Large Language Models (LLMs) to mine complex behavioral data, deriving deep, actionable shopper insights that identify customer experience gaps and unlock new business opportunities. * Design and develop scalable machine learning and GenAI models focused on shopper intent and preference modeling, ensuring a rapid path from prototype to production. * Partner closely with engineering teams to architect and deploy end-to-end GenAI solutions into production, integrating advanced insights directly into real-time, customer-facing systems. * Drive the scalability, efficiency, and automation of large-scale model training and real-time inference systems, pioneering the LLM infrastructure required to support next-generation GenAI workloads at Amazon Ads scale. * Design and run rigorous A/B experiments to quantify the business and customer impact of GenAI-driven shopper insights, performing advanced statistical analysis to guide iterative production rollouts. * Conduct applied research in novel generative AI techniques (e.g., fine-tuning, RAG, agentic workflows) to optimize the shopper experience and drive performance across all aspects of the Sponsored Products and Brands business.
Applied Scientist, Agentic WorkSpaces (AAWS)
US, WA, Seattle
AWS Applied AI Solutions (AAIS) is where science meets customer obsession at scale. We build the intelligent systems that power AWS services used by millions, combining research in machine learning, agentic AI, and applied science with the operational rigor required to deliver enterprise grade experiences. Within AAIS, Amazon WorkSpaces is our cloud based virtual desktop service that delivers secure, managed computing to over one million daily users across the globe, enabling organizations to provision, manage, and scale desktops with the reliability and performance their workforce depends on. We are looking for an Applied Scientist to be part of the tiger team building the capacity modelling for Amazon WorkSpaces, owning and advancing the science behind it. You will design, build, and continuously improve the forecasting and optimization models that ensure the right compute, storage, and networking resources are available at the right time, in the right regions, at the lowest possible cost, without ever compromising the end user experience. This is a high impact individual contributor role for someone who thrives at the intersection of applied research and production systems. You will define the scientific roadmap for capacity intelligence, turning reactive provisioning into a predictive, self optimizing engine that anticipates demand before customers feel any constraint. Key job responsibilities Define and drive the scientific strategy for capacity modelling, establishing the research agenda that transforms how WorkSpaces forecasts demand, plans supply, and allocates resources across a globally distributed infrastructure. Build advanced demand forecasting models that predict workspace usage across multiple time horizons, from intraday spikes to long range growth trajectories, incorporating signals such as customer onboarding patterns, seasonal trends, regional expansion, and macroeconomic indicators. Design supply optimization frameworks that determine optimal resource placement, instance mix, and pre warming strategies, balancing availability, performance, and cost by reasoning over hardware constraints, pricing dynamics, and service level objectives. Develop causal and probabilistic models that move beyond trend extrapolation to true understanding of demand drivers, enabling the organization to distinguish organic growth from one time events, anticipate shifts in usage patterns, and quantify uncertainty in planning decisions. Architect simulation and scenario planning systems that allow business and engineering leaders to run what if analyses, stress test capacity plans against disruption scenarios, and evaluate trade offs between investment timing, risk tolerance, and customer experience. Pioneer the integration of machine learning with operations research, combining deep learning based forecasting with mathematical optimization to jointly solve the demand prediction and resource allocation problem in a way that neither discipline can achieve alone. Establish evaluation frameworks and monitoring systems that measure forecast accuracy, capacity utilization, and cost efficiency in production, creating tight feedback loops that drive continuous model improvement and build organizational trust in science driven planning. Influence the broader organization's capacity strategy by translating model outputs into actionable recommendations for leadership, identifying opportunities to extend capacity intelligence patterns to adjacent services, and mentoring scientists and engineers across the team. About the team As part of the AWS solutions organization, we have a vision to provide business applications, leveraging Amazon's unique experience and expertise, that are used by millions of companies worldwide to manage day-to-day operations. We will accomplish this by accelerating our customers' businesses through delivery of intuitive and differentiated technology solutions that solve enduring business challenges. we blend vision with curiosity and Amazon's real-world experience to build opinionated, turnkey solutions. Where customers prefer to buy over build, we become their trusted partner with solutions that are no-brainers to buy and easy to use. Diverse Experiences AWS values diverse experiences. Even if you do not meet all of the preferred qualifications and skills listed in the job description, we encourage candidates to apply. If your career is just starting, hasn’t followed a traditional path, or includes alternative experiences, don’t let it stop you from applying. Why AWS? Amazon Web Services (AWS) is the world’s most comprehensive and broadly adopted cloud platform. We pioneered cloud computing and never stopped innovating — that’s why customers from the most successful startups to Global 500 companies trust our robust suite of products and services to power their businesses. Inclusive Team Culture AWS values curiosity and connection. Our employee-led and company-sponsored affinity groups promote inclusion and empower our people to take pride in what makes us unique. Our inclusion events foster stronger, more collaborative teams. Our continual innovation is fueled by the bold ideas, fresh perspectives, and passionate voices our teams bring to everything we do. Mentorship & Career Growth We’re continuously raising our performance bar as we strive to become Earth’s Best Employer. That’s why you’ll find endless knowledge-sharing, mentorship and other career-advancing resources here to help you develop into a better-rounded professional. Work/Life Balance We value work-life harmony. Achieving success at work should never come at the expense of sacrifices at home, which is why we strive for flexibility as part of our working culture. When we feel supported in the workplace and at home, there’s nothing we can’t achieve.
Applied Scientist, Demand Tech, Amazon Ads
GB, London
In Amazon Advertising, we apply machine learning at massive scale to optimize the prediction, ranking, and bidding behind every ad — deciding, in milliseconds, which ads to show shoppers and how to value them. We're looking for an Applied Scientist to help make sure the ads shoppers see are the right ones for them. You'll work across the science of how we rank, value, and bid on ads for Amazon DSP (Amazon's Demand-Side Platform) — including how we judge whether an ad is a good fit for the page a shopper is on and for the shopper themselves. It's high-scale, low-latency, customer-facing science: your models run live in front of millions of shoppers under tight real-time constraints. The questions are genuinely open — how do you tell whether an ad is relevant to someone, how do you balance what's good for shoppers, advertisers, and Amazon, and how do you keep getting that right as shopping behavior and inventory shift underneath you? Your work will have real impact, and you'll have room to shape where we take it. A few things make this stand out: your models touch a huge share of the ads shoppers see every day, so even small improvements add up fast; you'll run modern ML live under strict latency limits, across regions and very different types of ad inventory; and the problem space is rich — from how we value and bid on ads, to keeping models stable as traffic shifts, to what makes an ad a good fit for a shopper. Key job responsibilities - Design and improve the models that decide how ads are ranked, valued, and priced — including how relevant an ad is to the page and the shopper. - Apply and extend state-of-the-art techniques across e.g. ranking, deep learning, and information retrieval. - Own problems end to end: frame them, prototype, experiment, and ship them to production. - Balance competing objectives — shopper experience, advertiser and publisher value, and Amazon's business — into models that hold up across placements and marketplaces. - Communicate your work clearly to both business and science audiences, tailoring how you share it to each. - Write and ship your own production code backed by strong engineering support — we're all builders here. - Move fast with the best tools available, including modern AI coding assistants and agents. A day in the life You might start by digging into last week's experiment results, then use an AI coding agent to get your next prototype built and ready to test in production. In the afternoon you could be sketching a new way to measure ad relevance, reading a recent paper that bears on it, and talking it through with a senior scientist on the team. You'll move between hands-on science, writing and shipping real production code, and making the calls on your own work. About the team We're a group of scientists and engineers based in Edinburgh and London, working to make Amazon's ads more performant and relevant. We sit within a larger team spread primarily across New York City and the UK, and we have a broad mandate to build and experiment. You'll work alongside senior applied scientists you can learn from, with the data and infrastructure to do the work well and room to grow — with opportunities to attend top conferences (e.g., NeurIPS, KDD, ICML) and take on more scope over time.
Applied Scientist II, Amazon Shipping
IN, HR, Gurugram
Work on ML teams building large-scale forecasting and optimization systems that power Amazon’s global transportation network and directly impact customer experience and cost. As an Applied Scientist II, you will set scientific direction, mentor applied scientists, and partner with engineering and product leaders to deliver production-grade ML solutions at massive scale. Key job responsibilities 1. Lead and grow a high-performing team of Applied Scientists, providing technical guidance, mentorship, and career development. 2. Define and own the scientific vision and roadmap for ML solutions powering large-scale transportation planning and execution. 3. Guide model and system design across a range of techniques, including tree-based models, deep learning (LSTMs, transformers), LLMs, and reinforcement learning. 4. Ensure models are production-ready, scalable, and robust through close partnership with stakeholders. Partner with Product, Operations, and Engineering leaders to enable proactive decision-making and corrective actions. 5. Own end-to-end business metrics, directly influencing customer experience, cost optimization, and network reliability. 6. Help contribute to the broader ML community through publications, conference submissions, and internal knowledge sharing. A day in the life Your day includes reviewing model performance and business metrics, guiding technical design and experimentation, mentoring scientists, and driving roadmap execution. You’ll balance near-term delivery with long-term innovation while ensuring solutions are robust, interpretable, and scalable. Ultimately, your work helps improve delivery reliability, reduce costs, and enhance the customer experience at massive scale.
Principal Applied Scientist - AI for Life Sciences, AWS Applied AI Solutions - Life Sciences
US, WA, Seattle
As part of the AWS Applied AI Solutions organization, we have a vision to provide end user applications, leveraging Amazon's unique experience and expertise, that are used by millions of companies worldwide to manage day-to-day operations. We will accomplish this by accelerating our customers' businesses through delivery of intuitive and differentiated technology solutions that solve enduring business challenges. We blend vision with curiosity and Amazon's real-world experience to build opinionated, turnkey solutions. Where customers prefer to buy over build, we become their trusted partner with solutions that are easy to adopt and easy to use. The Team Join the next science revolution at AWS Life Sciences Applied AI Solutions, where you'll work alongside world-class scientists to build AI that transforms how therapeutics are discovered, developed, and brought to patients. We're out to revolutionize how medicines are discovered, developed, and brought to patients, powered by a new generation of AI. Our team tackles some of the hardest open problems at the intersection of frontier AI and life sciences. We apply biological foundation models, large language models, and agentic reasoning systems to life sciences problems, then put them into the hands of pharma, biotech, and diagnostics customers as applications and managed services they can fine-tune, tailor, and deploy on their own data. The science challenges are deep: how do you design agentic systems that reason correctly over complex biological, regulatory, and clinical logic? How do you enable customers to tailor foundation models to their proprietary data and get better outputs with less effort? How do you adapt models to reason faithfully in high-stakes scientific and regulatory domains? Today we're focused on two areas. In drug design, our products (including Amazon Bio Discovery) accelerate discovery by giving bench scientists AI-guided protein engineering and antibody design capabilities. In clinical trials, we're building AI that automates and optimizes regulatory and clinical development workflows. We combine frontier research with production-scale delivery to put breakthrough science into the hands of customers solving humanity's hardest problems. We value scientific rigor, encourage publication, and support conference participation. If you want to do research that ships, this is the team. The Role We are seeking an exceptional Principal Applied Scientist to set the scientific direction for our life sciences AI portfolio. You will be the scientific leader who defines research agendas, architects novel approaches, and delivers models and methods that give our customers capabilities that did not previously exist. This is a rare role that combines deep expertise in LLM reasoning and agentic AI with applied impact in life sciences. You will innovate on how large language models reason, plan, and act in complex scientific domains, while applying domain knowledge in biology to ensure models produce scientifically valid outputs. The problems span multiple fronts: - How do you build LLM-based agentic systems that correctly reason over clinical protocols, regulatory standards, and complex multi-step scientific workflows? - How do you develop model customization and training methods that let customers get state-of-the-art results from foundation models? - How do you adapt and extend protein and antibody models so customers can fine-tune on proprietary sequence data and get therapeutically relevant outputs? You will work across drug discovery (protein engineering, antibody design) and clinical trial operations (agentic automation, structured reasoning, domain adaptation). You will own end-to-end scientific solutions from research through production, and your work will directly shape the tools that thousands of scientists use daily. Key job responsibilities - Set the scientific vision and research agenda for LLM reasoning, agentic AI, and biological model customization across the portfolio - Innovate on LLM reasoning, planning, and agentic approaches for complex scientific and regulatory workflows - Develop model customization methods (fine-tuning, RLHF, retrieval augmentation, domain adaptation) that enable customers to train better models on their own data with less effort - Advance methods to adapt and extend biological foundation models for customer-specific therapeutic applications - Solve open research problems in faithful reasoning, multi-step planning, and tool use in high-stakes scientific domains - Partner with Life Sciences domain experts and customers to understand their hardest scientific challenges and translate those into tractable research problems - Publish at top-tier venues and build the team's external scientific reputation - Mentor applied scientists across the team while maintaining significant personal research contribution - Collaborate with product and engineering to ensure research translates into shipped products that serve customers at scale - Influence multi-year research roadmaps through deep scientific expertise and customer understanding A day in the life - Push a new reasoning approach into production that measurably improves outputs for a pharma customer's workflow - Design and run experiments to validate a novel fine-tuning method, then ship it as a capability customers can use immediately - Unblock a delivery milestone by diagnosing why a model is failing on a new class of inputs and implementing a fix - Meet with a customer's scientific team to scope what the next model release needs to do for them - Review a teammate's experimental results, sharpen the approach, and help get it over the finish line - Publish results from shipped work at a top venue, closing the loop between research and impact - Prototype a new idea that could become the next major capability in the product
Applied Scientist II, WW Prime - AI/ML Science
US, WA, Seattle
Interested in modeling and understanding customer behavior through machine learning, artificial intelligence, and data mining over TB scale data with huge business impact on millions of customers? Join our team of Scientists developing models to model customer behavior and optimize the customer experience with Amazon Prime. This includes understanding who our customers are, long-term value of the Prime membership program, and creating the right personalized framework for content and subscription optimization. As an AI/ML expert, you will partner directly with product owners to intake, build, and directly apply your modeling solutions. There are numerous scientific and technical challenges you will get to tackle in this role, such as optimizing/fine-tuning GenAI/LLM solutions for Prime personalization, building GenAI foundation models, global scalability of models, combinatorial optimization, cold start problem, accelerated experimentation, short/long term goals modeling, and multi-step optimization leading to reinforcement learning of the customer journey. We employ techniques from GenAI/LLMs, supervised/semi-supervised learning, deep learning, transformer architectures, using outcomes from causal Econometric modeling, and Reinforcement learning. As the central science team within Prime, our expertise gets routinely called upon to weigh in on a variety of topics. We also emphasize the need and value of scientific research and have developed a strong publication and patent record (internally/externally) which you will be a part of. You will also utilize and be exposed to the latest in ML technologies and infrastructure: AWS technologies (EMR/Spark, Sagemaker, DynamoDB, S3, ClaudeCode), various AI/ML algorithms and techniques (Deep Learning, GenAI/LLMs, transformers, supervised/unsupervised/semi-supervised/reinforcement learning), and statistical modeling techniques. - Stay abreast of current literature in the field and advance/build novel science solutions leveraging SoTA solutions. - Build and develop AI/ML models and supporting infrastructure at TB scale, in coordination with software engineering teams. - Leverage Deep Learning and GenAI solutions for building foundation models and personalized optimization solution. - Develop offline policy estimation tools and integrate with measurement systems/econometric models. - Establish scalable, efficient, automated processes for large scale data analyses, science development, science validation and model implementation. - Analyze and extract relevant information from large amounts of Amazon’s historical business data to help automate and optimize key processes. - Work closely with the business to understand their problem space, identify the opportunities and formulate the problems. - Use AI/machine learning, data mining, statistical techniques and others to create actionable, meaningful, and scalable solutions for the business problems. - Design, develop and evaluate highly innovative models and statistical approaches to understand and predict customer behavior and to solve business problems. Key job responsibilities Key job responsibilities - Stay abreast of current literature in the field and advance/build novel science solutions leveraging SoTA solutions. - Build and develop AI/ML models and supporting infrastructure at TB scale, in coordination with software engineering teams. - Leverage Deep Learning and GenAI solutions for building foundation models and personalized optimization solution. - Develop offline policy estimation tools and integrate with measurement systems/econometric models. - Establish scalable, efficient, automated processes for large scale data analyses, science development, science validation and model implementation. - Analyze and extract relevant information from large amounts of Amazon’s historical business data to help automate and optimize key processes. - Work closely with the business to understand their problem space, identify the opportunities and formulate the problems. - Use AI/machine learning, data mining, statistical techniques and others to create actionable, meaningful, and scalable solutions for the business problems. - Design, develop and evaluate highly innovative models and statistical approaches to understand and predict customer behavior and to solve business problems.
See more jobs
View from space of a connected network around planet Earth representing the Internet of Things.
Get more from Amazon Science
Subscribe to our newsletter

Amazon.com | Conditions of Use | Privacy | © 1996-2026 Amazon.com, Inc. or its affiliates

Social