For many of us, the use of our voices to interact with computers, phones and other devices is a relatively new experience made possible by services such as Amazon’s Alexa.
But it’s old hat for Luu Tra.
An Amazon Senior Main Engineer, Tran has spoken to computers for more than three decades. Tran a Uber-early adopter of voice computing, Tran remembers the days when PCs came without sound cards, microphones or even sound jack. So he built his own solution.
“I remember when I got my first Sound Blaster Sound card that came with a microphone and software called Dragon Natural Speaking,” Tran remembers.
With a little plug-and-palay technique, TRA could suddenly use his voice to open and save files in the mid-1990s PC. Replacing his keyboard and mouse with his voice was a magical experience and gave him a glump into the future of voice -controlled computing.
Frop until 2023, and we at the Golden Age of Voice Computing, made possible by progress in machine learning, AI and voice assistants like Alexa. “Amazon’s vision for Alexa was always to be a conversation, natural personal assistant who knows you, understands you and has some personality,” says Tran.
In its role, Tran has monitored the planning-Deploy-scale cycle for many Alexa features: timers, alarms, reminders, calendar, recipes, falls, messages and more. Now he helps Amazon by facilitating collaboration between the company’s engineers and academic researchers who can help promote machine learning and Ai-Both full-time academics and those who participate in Amazon’s scholars and visit academic programs.
Tran is no stranger to the calculation of paradigm shifts. His past experiences at Akamai, Mint.com and Intuit gave him a front-drawn seat to some of Tech’s most dramatic shifts, including the birth of the Internet, the explosion of mobile and the shift from computer computing.
Bringing his three decades of experience to bear in his role with Amazon helps Tran explore the potential of voice competence by spuring collaborations between Amazon’s Technology and Science Teams. We have daily, Tran encourages engineers and scientists to work together as a one-shoulder-to-store-merge the latest scientific research with advanced technique.
It is no crane accident is to help lead Alexa’s next engineering chapter. To grow up by looking at Star TrekHe had always been fascinated by the idea that you could talk to the computer and it could speak back using AI.
“I had always thought that AI was out of reach of my career and longevity. But now look at where we Teday,” says Tran.
The science of technique Alexa
Tran believes that collaboration with researchers is important to continue innovation, both with Alexa and AI in general.
I come from the perspective of an engineer who has studied some theory but has worked for decades translated technological ideas into reality, within the constraints of the real world.
“Bringing them together – engineering and the science – is a strong combination. Many of our projects are not simple deterministic technical problems we can solve with more codes and better algorithms,” he says. “We need to bring a lot of different technical and leverage science to fill the gaps, such as machine learning modeling and training.”
Helping engineers and researchers work closely together is a non -trivial effort because they often come from different backgrounds, have different goals and incentives and in some cases even speak different “language.” E.g. Points to Tran that the word “function” means something very different from product conductors and engineers than for researchers.
“I come from the perspective of an engineer who has studied some theory, but has worked for decades translated technological ideas into reality in the realm of the real world. For me it has been important to understand why Something works than what Works, ”says Tran.
To realize the best of both worlds, says Tran, says the Alexa team a flexible approach than used in the past – to bring together project teams by product leaders, engineers and researchers, often with different combinations based on the required goal or technically required. There is no dogma or doctrine that says what roles should be on a particular team.
What is the most is that Tran points out is that each team from the start of the customer needing, the use case, the product market and even the monetization strategy. Bringing researchers into projects from the start is critical. “We always have product leaders on teams with engineers and scientists. Some teams are divided 50-50 between researchers and engineers. Some are 90% researchers. It just depends on the problem we are looking for.”
Makeup of teams changes as the projects progress. Some start heavily weighted against technique and then determine a use or problem that research research. Others start predominantly science -based, and when a viable solution is in sight, several engineers gradually add to building, testing and iterers. This push/drag among how teams form and change – and the car to organize and reorganize quickly to iterate – is key, says Tran.
“Often it is still product conductors who describe the core customer need and use and how we go to solve it,” says Tran. “Then the researchers will say, ‘yes, it’s stupid or no, it’s still science fiction.’ And then we iter and formalize the kind of project.
Engineering + Science = smarter recipe recommendations
A recent project that is benefited from the new agile, collaboration method, is Alexa’s new contradiction engine. To deliver a receiving recipient recommendation to a customer asking for a – perhaps for an Amazon Echo show on a kitchen table – Alexa has to choose a single recipes from her huge collection while understanding the customer’s wishes and context. All of us have unique taste, dietary preferences, potential food allergies and contextual factors in real time, such as what is in the fridge, what time of day it is and how much time we have to cook a meal.
This is not something you can build using brute engineering, it takes a lot of science.
Alexa, explains Tran, Factor must all parameters in recipe recommendation and – in milliseconds – return a recipe that it thinks is both very high (such as a Mexican dish) and personal (eg no meat for vegetarian customers). The technology involved to respond within, secure, satisfactory recommendations to any customer is thoughtful-bally complex. “This is not something you can build the use of brute-fun technique,” Tran notes. “It takes a lot of science.”
Building the new recipe engine required two parallel projects: a new machine learning model trained to look through and choose recipes from a corpus of millions of online recipes and a new inference engine to ensure that each request Alexa receiver is called with identified personal and contextual data. “We broke it down, just like any other process of building software,” says Tran. “We wrote our plan, identified the tasks and then decided that each task was best handled by a scientist or an engineer, or perhaps a combination of both working together.
Tran says the researchers on the team largely focused on the machine learning model. They started by examining all existing, publicly available ML approaches to recipe recommendation – cataloging of the model types and narrowing of them based on what thought would work best. “The researchers looked at a lot of different approaches-Bayesian models, graph-based models, cross-cutting domain models, neural networks and collaborative filtering and settling on a set of six models they thought would be best for us,” explains Tran. “It quickly helped us to narrow without having to exhausting any potential model approach.”
The engineers, meanwhile, worked on designing and building the new inference engine to better capture and analyze user signals, both implicitly (eg time of day) and explicit (whether the user asked for a dinner or lunch recipe). “You would not recommend cocktail recipes at breakfast time, goals for people Lack Eating pancakes for dinner, ”haunts crannies.
The inference motor had to be built to meet queries from existing users and new users who have never asked for a recommendation recommendation. Performance and privacy were the most important requirements. The engineering team had to design and implement the engine to optimize the flow, while the minimization of calculation and storage costs and complexing with customer requests to delete personal information from their history.
When the new inference engine was clear, the engineers integrated it with the six ML models built and trained by the researchers, connected with the new front-end interface built by the design team and tested the models against each other to compare the results. Tran says that all six models improved conversion (a “conversion event” is triggered when they need to use choose a recommended recipes) vs. Baseline recommendations, but a model surpasses others by more than 100%. The team chose the model that is in production today.
However, the write -up project does not end here. Now that it is live and in production, there is a process of continuous improvement. “We always learn from customer behavior. Which are the recipes that customers were really happy with? And which are they they never choose?” Says Tran. “There is still collaboration between engineers and researchers about it to refine the solution as well.”
Future: Alexa Engineering driven by science
To further accelerate Alexa Innovation, Amazon Alexa Main formed Community-A Matrixed Team of several chased engineers and researchers working on, and Alexa and Alexa-related technologies. “We have people from all parts of the company who look out for who they are reporting,” Tran adds. “What brings us together is that we worked together on the technologies behind Alexa, which is fantastic.”
Earlier this year, more than 100 members of this society, both personal and remote, called in to share, disk and debate Alexa technology. “In my role as a member of the small leadership team of the community, I presented a few sessions, but I was mostly to learn from, get in touch with and influence my peers.”
Tran thoroughly enjoys his work with researchers and he feels he is taking great advantage of the collaboration. “Working closely with many scientists helps me understand what that-art AI is capable of, so I can leave it in the system that I design and build. But they also help me understand its limitations so I don’t donate and try to build some just can’t be achieved in any realistic timeframe.”
Tran says that today more than ever is a great time to be at Alexa. “Fantasy has been locked up in the population and in our customer base,” he says. “So the next question they have is, ‘Where is Alexa going?’ And we were worn as we can to bridge new features for the life of customers.