Rajen Sheth, Google
Rajen Sheth is Senior Director, Product Management currently responsible for running the Google Cloud Artificial Intelligence and Machine Learning product line. This includes products which bring Google’s AI technology to enterprise customers and developers, including Google’s AI platform, image, video, and speech AI capabilities.
Previously, Rajen was responsible for the Android and Chrome for Business and Education group, leading a cross functional team across product, engineering, sales, support, ops, and partnerships. Earlier, Rajen also co-founded and led the Google Apps product line, and Google’s messaging and collaboration products for businesses. He was also responsible for leading Google’s efforts to bring cloud computing to universities and enterprises.
Rajen holds a B.S. and M.S. in computer science from Stanford University. Capgemini’s Digital Transformation Institute spoke to Rajen to understand how Google is democratizing AI.
Can you tell us a bit about your role at Google?
I’m the Director for our Google Cloud artificial intelligence team, where we bring Google’s many AI technologies to bear for our customers and for developers. We do a number of things. At the infrastructure level, we bring the best of infrastructure for artificial intelligence to customers. At the developer level, we give developers access to a lot of the great models that we’ve developed for products as well as those built specifically for enterprise use cases. Thirdly, we provide solutions that meet specific business problems.
AI is going to be in every industry in ten years
What are some of the implications of AI for large companies?
I think the industry is still figuring out where AI can be most beneficial for companies. But I do see a number of areas. For example, we are seeing a particular interest and impact in personalizing customer interaction. It is always hard to give a customer a personalized experience when you have multiple customers that are coming to your website. AI can help give a much more personalized interaction -everything from recommendations to helping interact with the customer via customer service applications such as chatbots. AI can help in contextual impact, so that the customer can get the answer they need quickly. AI helps in marketing campaigns and loyalty programs as well. Another area is efficiency and agility. Spotting new patterns in a database can lead to strong efficiencies. Another area is adding structure to unstructured data.
Which sector do you believe will be most affected in the short term?
We are definitely seeing retail as an area where there is a lot of potential for AI. Everything from how you personalize a customer interaction all the way through to how do rethink the concept of the store. For example, when you walk into a retail store, you are almost always anonymous. There’s barely any of the personalization you might have if you go to the brand’s website. There are ways that AI can help solve that and make the interaction with the customer a lot more personal. Within the next five to 10 years, almost every company will be using AI in a pretty significant way.
You don’t need to know AI to benefit from it
For an organization to benefit from AI, where should it begin?
I think it should begin with the business problem. It should not start with the technology. It really starts with figuring out what business problem you want to solve and then figuring out how AI can be applied.
If top AI talent is going to companies like Google, how should large organizations meet the AI talent challenge?
Talent is actually the biggest limiting factor for AI right now. There are very few people that can actually build a deep learning model for example. I think it is a two-sided challenge. On one hand, how do you train more people to be able to do these things? On the other hand, how do we make AI easy and useful for organizations? This is where we are focusing. How do we make it so that the existing teams in organizations can leverage AI? Existing developers in traditional organizations shouldn’t have to immerse themselves into the deep learning model in order to benefit from it.
AI today is really where the web and the Internet were in 1994. Everybody sees a lot of promise, but it is still very hard to build upon. Back then, building a web page required a lot of know-how in terms of knowing HTML, for example. From there, we very quickly got to the point where any average business owner could put up a website and enable e-commerce. And that’s where we need to get to with AI. The average business owner should have building blocks that they can put together in an easy way to be able to accomplish their needs without having to build models themselves. That is what we are focusing on – how do we bring AI to many more people, how do we democratize it?
AI still has its issues
What is your view on the black box nature of AI?
In many cases it doesn’t necessarily matter how AI came to a conclusion. The bigger question is about whether the conclusion is useful and ends up creating benefits? Take, for example, product recommendation. If AI recommends a particular product, the way you really test that is whether people buy the recommended product. Of course, there are many other areas where explainability is critical. For example, with medicine or fraud detection it is critical to be able to explain AI’s decisions. That is something where the science needs to evolve a little bit more because these algorithms are fairly generic. Figuring out explainability is a challenge that the field needs to work out.
In your view, what are some of the issues with potential biases resulting from the data that is used to train AI?
That is a big issue. AI is fundamentally not biased, but if the data itself is biased, then it is going to produce biased outputs. That is an area where we as a field need to improve and figure out how we get to the right kinds of best practices. One thing that we observe is that the vast majority of work in building a great model is not actually in the model itself, it is more in the data. Figuring out what data to use, getting the data to the right place, cleaning the data, doing data engineering and feature engineering, and then figuring out what outputs you are looking for and how to use those outputs. That is the bigger challenge in almost any of these problems.
AI @ Google
Google has a variety of AI initiatives. How does the company view AI?
At Google, we emphasize that AI is not something that should be done by only one group. AI is something that should be pervasive. Over the last two years, we’ve gone from around a couple hundred projects at Google using AI to almost 7,500 projects. Sundar [Sundar Pichai, Google CEO] has already said we are an AI-first company. We’ve gone through a rapid transition to becoming an AI company where almost every product utilizes AI as a tool to provide a better user experience. The current level of interest in AI is like nothing I’ve seen before in my career.
Which Google product do you believe AI will impact the most?
AI can have an impact on every Google product. I see AI as more of a paradigm of engineering as opposed to anything else. One of the best examples of AI at Google is Google Photos. All of us have tens of thousands of photos and categorizing them is a painful exercise. Google Photos really took that to the next level by bringing Google-quality search to photos. Translation is another big area. We just publicly shared a demo with the new Pixel 2 smartphone where, if you wear our ear buds, we can translate another speaker using a different language and give you the translation in your ear simultaneously. It is really magical because we can start to bring the world together. Another surprising application of AI for us has been the Gmail smart reply. With email on mobile, we suggest a reply at the bottom of each message. On mobile, that is extremely valuable, because typing something on a mobile phone is not an easy thing to do. We are now finding that over 10% of replies via mobile for Gmail use this smart request feature. Things like that are really showing that AI can not only be accurate but incredibly useful too.
How does AI help you differentiate Google Cloud?
AI is going to be a big differentiator for Google cloud in a number of ways. First is the infrastructure and making Google cloud the fastest place to run AI. Second is the building blocks – the prebuilt models that users can take and customize. We have a wealth of models and knowledge that are there for users to take and use. Third is partners: having partners build on top of this and offer services to customers is really how this can scale. We are putting a lot of emphasis on working closely with partners to make them successful.
What do you say to people that are worried about AI taking over the world?
AI is still a very nascent technology. We are still only a few years beyond being able to take an image and tell a dog from a cat. We still need to figure out how AI can best augment what we do. The big issue for me is one we spoke of earlier – bias. I think that is a more pressing issue than worries about taking over the world. If we are producing algorithms that contain biases, then that becomes a much bigger problem.
We are really at the beginning of a journey for something that is going to have a positive impact for people everywhere over the course of the next 10 to 20 years. It is up to us to figure out how best it will have an impact. I’m really excited about having this opportunity to really shape the next generation of technology.