Machine learning in the cloud

Marius Mårnes Mathiesen
Marius Mårnes Mathiesen | 3 April, 2019

Remember ten years ago, when everybody was talking about going mobile first? When the mobile web started gaining traction, quickly followed by the advent of native apps on smart phones, the IT business moved from viewing mobile users as a potentially interesting niche of users to becoming the main target for digital products and services. This trend – mobile first – has been a key driver in Shortcut’s success in the market. When we approached customers, talking about mobile apps around 2010, most organizations were skeptical to investing in mobile services, since they were still seeing far higher numbers of users on desktop platforms than on mobile. Since then, some of the largest new digital services have been mobile only, with apps like Snapchat, Instagram, Vipps and REMA 1000’s Æ as some huge digital success stories with a practically non-existant presence on the old (ie. desktop) web.

 

It should come as no surprise that many customers were just as skeptical when Google announced that they were moving from a mobile first to an AI first approach to digital product development last year. Artificial Intelligence has been a much researched topic within the industry for more than five decades, so why this change in focus today?

Although the theoretical research and actual algorithms which lies behind the broad term AI haven’t changed that much in the last 10 years, the feasability of actually applying techniques like machine learning has become much higher. The quality of Google’s image recognition technology, the appearing of cars with (limited) self driving abilities, and the computer program AlphaGo beating a professional player in the game of Go in 2015 are all examples where applied machine learning seems to have become so promising that it’s impossible to ignore. Although an AI passing the Turing Test could still be a few years into the future, the Intelligence part of Artificial Intelligence is improving rapidly and quite a few organizations are placing their bets on digital assistants being able to replace actual humans within a fairly short timeframe.

Why now?

So if all the algorithms and theory behind AI has existed for years, why is everyone talking about AI today?

One thing is that our industry has a tradition of exaggerating technology, which causes organizations to give too much focus to new technology, which in turn is followed by skepticism. This hype circle has not just caused organizations to waste resources on immature technology, it has also made them overly skeptical to the potential impact of technology on their business.

Another, in my opinion more important, factor behind the increased attention to AI is how applying machine learning to huge amounts of data has become feasible for more companies by the availability of vast processing power in the cloud. A lot of attention was given to being able to store and analyze big datasets (aka. Big Data) a few years back, and this was enabled by the cloud transformation happening in the industry. These days, the ability to not just analyze this data sets but also to make predictions based on this data using machine learning. In addition to the algorithms and the compute power to process enormous amounts of data, we are also seeing better tools for applying machine learning without investing huge amounts in low level model development. For example, doing object recognition in photos is possible even in real time on smart phones these days, with previously unthinkable precision. In the same manner, understanding natural language – including sentiment analysis and intent recognition – is possible using off the shelf software, powered by the cloud.

In addition to the availability of proven machine learning algorithms and vast amounts of processing capacity, the amount of data available today is also a key enabler for modern AI. In order to get good results from a machine learning model, huge amounts of data is required, and this has only recently become available after the digitization of business interactions. Imagine how for instance retailers today possess data on every single product sold in their stores over the last decade, along with data about the customers purchasing these products. By combining these data sets and building machine learning models, a retailer is able to make really good predictions on a customer’s next purchase or prevent churn.

Some of the most researched machine learning problems today can be applied in very short term using higher level products from Google:

AutoML Natural Language is a cloud hosted service for classifying natural langue text into labels. The cloud hosted service takes away some of the more tedious, low-level tasks of NLP by allowing you to upload a CSV file (ie. an Excel spreadsheet) containing text in one column and a label in the other column. By uploading a CSV file containing millions of example data, the service will build and train a model which after a few hours of data crunching on Google’s servers will be able to predict a label from text sent to the service. A use case for this could be to import real questions asked by a user in a support system and label this with a product or solution relevant for the user’s question. This way a business could build a prediction system for customer support requests and provide self service options provided by the business. And all this without attracting and training a machine learning expert to build a custom model for your business. We believe that most organizations who have started machine learning projects from scratch have experienced that this is both expensive, time consuming and often results in projects which fail to deliver value.

Cloud AutoML Vision is the same technology which powers Google Photos, but intended for recognizing the contents of images relevant to your business domain. With AutoML Vision a doctor could upload x-ray images of pasients with healthy lungs and lung cancer, and then use a proprietary model for recognizing tumors in pasients. Or a fashion store recommending clothes to a customer based on outfits they like.

BigQuery ML is machine learning technology built on top of Google’s BigQuery data warehouse. Where BigQuery easily aggregates huge datasets from various legacy and third party systems and exposes these data using SQL, BigQuery ML lets a company automatically build machine learning models based on existing data and run predictions on incoming data. On one of Google’s samples for BigQuery ML they use data on every child birth in the US over the last 40 years to predict birth weight on newborns today. This model predicts birth date based on the baby’s gender, the length of the pregnancy, and demographic information about the mother - all of which have impact on the birth weight of the child. What really sets BigQuery ML apart from other solutions to the same problem is that no machine learning experience is required to use it; rather the well-known SQL language is used to build models.

How to get started with high impact machine learning projects

Many of the companies we work with at Shortcut have already come far in harvesting business data, identifying potential use cases and even run proofs of concept projects using machine learning. However, not many companies have gotten to a point where AI is actually generating results in their business. One problem we see is that the quality of data in these organizations are lacking, another one is that the organization simply doesn’t have the bandwith or capacity to take on new projects using their data.

We believe that machine learning projects need to bring various parts of the organization into the same room to identify opportunities and work closely together to realize these opportunities. And although working with machine learning is a process where it is difficult to predict the outcomes, we think it’s important to get valuable results as quickly as possible.

As a Google Partner, Shortcut has received specialist training from Google’s services organization PSO. Based on experience from working with machine learning inside of Google and for large corporations in various industries across the world, they have built a methodology and training materials which they train selected partners in so their partners can apply these learnings and techniques for their clients.

If your organization wants to use Google’s experience and approach to applying machine learning to your business, please get in with us to get the ball rolling. One example of offerings we can help you with to get started is an offering called Cloud Discover: Machine Learning where we over a week help your business get started with machine learning project which should provide business value within 1-3 months.

Please get in touch with us to start planning for your transistion to an AI company.

 

Topics

machine-learning googlecloud