Should Google be your AI and machine learning platform?

There's an arms race among public cloud providers to build the best machine learning platform and capabilities. Here's a look at what Google brings to the table.
Written by Natalie Gagliordi, Contributor
Image: Google

In a little less than 20 years, Google has evolved from a search engine experiment at Stanford University to a technology behemoth that's synonymous with the discovery of information.

Many of Google's core services straddle the business and consumer worlds, and as such serve as the backdrop for many of the menial, digital tasks we execute on a daily basis. But despite its relative ubiquity, there are still pivotal areas in tech where Google faces fierce competition.

Machine learning, the concept of training large-scale AI networks to teach and improve themselves over time, is one such area. There's an arms race among public cloud providers to build the best machine learning capabilities for enterprises interested in creating their own intelligent applications. Here's a look at what Google brings to the table.

The backstory

"Machine learning has been infused into Google products and services for over a decade," said Rob Craft, Google Cloud's product lead for Cloud Machine Learning. "Google first began to use machine learning by applying it to our products that serve billions of users."

Craft said Google's dive into AI and machine learning began as a matter of necessity after the company reached a scalability high mark with its original rules-based programming system. It was then that Google shifted to a learning system that would expand the boundaries of its entire platform.

"Rules are fragile. We built a great business using rules, but at some point we couldn't use it anymore," Craft said, pointing out that two of Google's legacy products -- Search and Maps -- were among the earliest beneficiaries of Google's machine learning advances.

Today machine learning touches nearly every Google product. For instance, the Google app uses speech recognition and natural language processing to understand speech in 55 languages; Google Search uses RankBrain, a ranking signal that uses deep learning to improve results; and Google Photos taps into the company's latest image recognition system.

Internally, the push into machine learning prompted the formation of dedicated teams that are actively using the technology to improve Google's consumer products, cloud platform, and Google's own business.

Google Cloud, Google's enterprise division headed up by Diane Greene, employs dedicated engineering and product teams to create and build machine learning tools and services, Craft said. More recently, Google announced the creation of a new machine learning team -- helmed by Fei-Fei Li, formerly the director of AI at Stanford, and Jia Li, who was previously head of research at Snap, Inc -- as part of an effort to unify some of the disparate machine learning work across Google's cloud.


Google's machine learning portfolio includes a range of cloud services and tools. In 2015, Google helped accelerate adoption of its services by open-sourcing its proprietary machine-learning library TensorFlow. Craft said TensorFlow is now the most popular machine learning on the software repository GitHub, with contributions coming from mostly outside Google.

As a follow up, Google Cloud Platform released a dedicated set of machine learning APIs based on Google created pre-trained machine learning models, which Craft said are meant to give developers access to high-quality cognitive services. The API set includes the Translation, Cloud Vision, Natural Language, Speech and Jobs APIs.

Cloud Machine Learning is Google's fully managed service that lets users create neural network and algorithm models and also run predictions at scale without worrying about the infrastructure. The service utilizes several of Google Cloud's data analytic tools such as BigQuery, DataFlow, and Datalab.

Google is also leveraging machine learning to power its own infrastructure that's used by Google Cloud users.

"For example, in our data centers, we're using machine learning to reduce the amount of electricity needed for cooling by 40 percent," Craft said. "This results in cost benefits to our users, as well as a greener planet."

One of the more prominent enterprises using Google's machine learning platform is Evernote. The note-taking service announced it was moving its infrastructure to the Google Cloud Platform in September and, according to CTO Anirban Kundu, the results have exceeded expectations. Kundu said that Evernote is now using Google's speech-to-text service, Translate API, Google Natural Language API, and managed machine learning.

"The power of machine learning and AI is that it will help Evernote users not only to remember everything but also to turn those ideas into actions, to actually help them think," Kundu said.

Challenges and ethics

While most machine-learning vendors will cite technical hurdles related to development of AI and machine learning systems, Craft noted a more nuanced challenge Google experienced.

In the case of Google's search team, Craft said there was a short learning curve when the team figured out that learning systems could be used to improve search functionality with just a few months of work.

"The difficulty was convincing the team to take consultations with a machine learning expert -- it was a culture aspect," Craft said. "The technology alone was one part. But search was the marquee service of Google, there was a culture there."

In terms of ethical challenges, Google is among the more prominent voices pushing for AI safety research and conversations surrounding ethics. Most recently, Google announced a partnership with Amazon, Facebook, IBM, and Microsoft to advance public knowledge of AI and formulate best practices. The tech giant also released new research on AI Safety in collaboration with OpenAI, Stanford, and Berkeley.

"We believe AI will be overwhelmingly beneficial -- we already see how machine learning is improving people's lives. We want to ensure machine learning is useful and helpful for everyone and we'll continue to do more research," Craft said. "This is an important aspect as the industry collectively develops technology in a responsible manner."

Advantages and vision

According to Craft, the strengths of Google's machine learning platform can be broken down into two main categories: Building the broader research community around machine learning and democratizing machine learning tools and services for businesses.

"Google is heavily invested in growing the research community around machine learning and technology," Craft said. "Fostering a broader community around machine learning will accelerate overall adoption and breakthroughs in its potential to change the world."

This open-source strategy is closely related to Google's decision to use its public cloud infrastructure for machine learning, which has helped drive down costs and requirements for infrastructure expertise while also boosting accessibility.

"By using the public cloud, this allows us to offer the latest services and advancements in technology for our users such as fluidly adding improvements to our existing offerings, spinning up new services at a faster rate, and supporting with the necessary infrastructure at no additional cost to our users," Craft said.

Another way to look at Google's machine learning prowess is to consider the tremendous amount of data and compute power at its disposal.

"Data and compute infrastructure are the makings of machine learning and AI," said Alexander Linden, research VP for machine learning and data science at Gartner. "They are the essence that brings Google at the forefront there."

Looking ahead, Google's vision for machine learning and AI is focused on the positive societal and economical changes that will come with its pervasiveness.

"Machine learning is going to drive the next wave in cloud computing," said Craft. "We foresee a future where machine learning delivers benefits across many different industries. The next step is to create deeper and customized machine learning models for specific industry use-cases, and we anticipate to continue to build our portfolio and services for everyone."

Also see:

Editorial standards