Uber is really one big Machine Learning problem – determining a route, how much to charge, how to pair riders in Pool – and AI Labs helps find solutions to all these problems and more across the many organizations at Uber. Did you know that the Uber AI group is made up of 5 different teams and 96 employees working on supporting Artificial Intelligence Research for Uber globally? Led by Chief Scientist Zoubin Ghahramani, Uber AI is doing exciting work in areas of research including Natural Language Processing, Bayesian Optimization, Neuroevolution, Reinforcement Learning, Deep Learning, and Computer Vision.
What attracted you to Uber?
“Our team, Conversational AI (Artificial Intelligence), is situated between research and product. I get the best of both worlds – interesting new problems and I get to see the results by releasing the solutions to real users. We’re working on a Hands-Free Voice Assistant for drivers. A voice assistant for drivers is a no brainer when it comes to safety. Drivers are already multi-tasking: watching out for traffic, looking out for pedestrians, looking for their rider and the last thing we want is for them to have to touch the screen any more than is necessary or legal. So our product will enable them to accomplish critical tasks via their voice.”
Describe what your team is responsible for?
“Conversational AI is an interdisciplinary group of researchers, engineers, and data scientists. I support our team by crowdsourcing and engineering raw data prior to a product release, adding meaningful structure to the data that an AI system can learn from. I do data analysis and metrics on natural language data to evaluate model performance. I absolutely love it when scientists come to me with an idea for a model. I get to help them think through problems and come up with a solution. Basically, I demonstrate how complicated language is and then I help untangle how meaning is constructed. We work together to find a solution that captures a human’s understanding of language in that context, is learnable by models, and works for our product.”
What are the biggest challenges you face in this role?
“Unlike a lot of other AI that needs human-labeled data and never changes, our understanding of language changes all the time. We may change the context that sentence will be used. We may add scenarios that create a category that overlaps with another category. We may change the business logic on how we want to handle ambiguity and scenarios not currently covered. One annotator may interpret the phrase one way while another annotator interpreted it differently. We need to update and re-interpret (re-label) natural language data often. Adding full dialog contextual understanding (handling a longer back and forth conversation between the driver and the assistant) will add another layer of labeling complexity.
Dialog context management in the context of Uber is particularly interesting since an AI assistant will have to handle frequent context switching, deciding which trip or rider the driver is replying to or whether the driver is referring to another voice-enabled task. The system could potentially handle a conversation between three participants in the conversation: the rider, driver, and the AI assistant. Three participants adds another layer of semantic complexity beyond the usual two participants (user~system) conversation that we currently see in products. You need a data pipeline that is compliant to GDPR and other regulations, preserving privacy, and establishing the right principles from day one and can handle the complexity and frequent updates. Creating a data pipeline isn’t sexy work, but it can make or break how quickly and iteratively you can ship high-quality models.”
Apply for a role on the Team