In June 2022, Amazon re:MARS, the company’s in-person event that explores advancements and practical applications within machine learning, automation, robotics, and space (MARS), took place in Las Vegas. The event brought together thought leaders and technical experts building the future of artificial intelligence and machine learning, and included keynote talks, innovation spotlights, and a series of breakout-session talks.
Now, in our re:MARS revisited series, Amazon Science is taking a look back at some of the keynotes, and breakout session talks from the conference. We've asked presenters three questions about their talks, and provide the full video of their presentation.
On June 27, Alexa Speech employees Aaron Eakin, senior principal engineer, and Angela Sun, senior manager, product management, presented their talk, "Self-learning Alexa: ML model updates with no human in the loop". Their talk focused on a real-time continual, lifelong learning system that trains machine learning models using production data at scale, without persisting data or retraining from scratch.
What was the central theme of your presentation?
Improving machine learning models via self-learning approaches without human labeling. This accelerates the rate of improvement from days/weeks/months to near real-time, which enables us to keep up with trending topics and shifting usage patterns, while also protecting customer privacy.
In what applications do you expect this work to have the biggest impact?
Applications that use neural-based machine learning models, where the data distribution shifts rapidly over time and/or where the data is privacy-sensitive.
What are the key points you hope audiences take away from your talk?
- Self-learning approaches can be used successfully to replace human labeling in certain models, as shown in our automatic speech recognition system.
- These approaches can enable models to effectively keep up with shifting usage patterns over time.
- The near-real-time approach of this system can also help improve customer privacy by not requiring data to be persisted or reviewed by humans.
