Machine Learning Predictive APIs and Apps: a first-hand report from the recent second International Conference in Sydney, Australia
Last week I had the good fortune to be able to attend the Second International Conference on PAPIs held in Sydney. Let me share a brief summary of all the presentations that were held on the first day and try to highlight some of the more interesting machine learning facts, tricks, and trends that came my way.
Real-World Predictive Applications
The Keynote address was on Real-World Predictive Applications with Amazon Machine Learning and was delivered by Danny Lange himself – the General Manager of Amazon Machine Learning. Danny began by saying that building smart apps is hard and by the end of the day I was definitely ready to agree with him. He also stated that there is a dearth of Data Scientists available to fill the needed positions in this field. So if you love data and wouldn’t mind being employed in the near future, you should probably set your sites on studying up on Predictive APIs.
To be honest, I didn’t gain too much from the first part of the talk – it was mostly an overview of Amazon Machine Learning.
However, his real-world example was fascinating: an end-to-end social media listening application powered by ML technology.
The application continuously monitors all tweets that mention your company’s Twitter handle, and predicts whether or not your company’s customer support team should reach out to the poster. Danny showed how, by using a Machine Learning model as your first tier of support, you can lower support costs and increase customer satisfaction. A simple but effective example of Amazon Machine Learning in action. He also showed how the application integrates Amazon Machine Learning with Amazon Mechanical Turk, Amazon Kinesis, AWS Lambda, and Amazon Simple Notification Service (Amazon SNS).
Cloud Academy now has a full course on Amazon Machine Learning to guide you through the whole process, from working with large data sources to generating accurate predictions.
Small Data is the new Big Data
Next up was the Technical Director and Partner at Resolve Digital, David Jones. His basic theme was that small data can be enough and that quality can be more important than blindly insisting on the more trendy Big Data. David then went on to show that you may not need as much data as you think to get real-world benefits.
David’s company procured a contract with an online wine retailer and, using off the shelf, well-implemented algorithms, generated a 71% increase in revenue. They didn’t use Google or AWS but rather an open-source system (whose name escapes me for the moment). Using open-source also helped keep costs down. David’s last word was that PAPI’s are awesome and that Machine Learning can be very effective when used in an eCommerce environment. I’ll drink to that.
Making Machine Learning accessible
Following David was Konstantin Davydov, a Software Engineer at Google who waxed lyrical about Simple Machine Learning for the masses. He said that, by using Google’s Cloud Machine Learning Services, users can set up an entire Machine Learning pipeline quickly and with limited or no Machine Learning expertise, and that it is also possible to build applications on top of the Prediction API that allows non-technical users to leverage the power of Machine Learning to help solve real-world problems.
I never thought that I was using Machine Learning in my day-to-day activities, but when Konstantin went on to explain, as an example, that the smart autofill add-on in Google Sheets is actually a predictive API, it dawned on me that Predictive APIs are probably more ubiquitous then I had thought.
Consumable, programmable, and scalable machine learning
After lunch, Poul E. J. Petersen, Chief Infrastructure Officer at BigML, hit the stage with what seemed to be the most technically detailed discussion of the day. He started off with a very optimistic statement:
In this tutorial you will learn how to perform classification, clustering, and anomaly detection tasks – all using the BigML REST API. By the end of this presentation, you will be ready to code your own predictive application in python using BigML.
Sorry to disappoint you Poul, but I can’t say that I’m quite there yet. Not being a python programmer I did find this presentation a bit of a strain, and the fact that it was straight after lunch didn’t help either: Poul shared a story from his youth where he would always fall asleep in Spanish class and thought he hated Spanish, but then realized it was the fact that the class was always after the lunch break, and in fact, he did like Spanish. So maybe I can code my own predictive application in python using BigML, but not straight after lunch.
To summarize Poul’s talk, he described how his team has been working hard for the past four years to democratize machine learning – making it more consumable, programmable, and scalable. Because of their well-defined workflow and powerful visualizations, it’s relatively easy for anyone to rapidly prototype ML solutions, and that at its core, BigML is really nothing more than a powerful and extensible Machine Learning API. In his tutorial Poul then showed how to perform classification, clustering, and anomaly detection tasks – all using the BigML REST API.
Predictive maintenance applications
Finally, Yan Zhang, a Microsoft data scientist showed us what’s available from Redmond. There were a few giggles from the audience when her Microsoft Windows Laptop failed to connect to the wi-fi. But to her credit, she recovered nicely and gave a very solid description of the landscape and challenges of predictive maintenance applications in the industry. She also did a good job illustrating how to formulate a problem with three different machine learning models (regression, binary classification, and multi-class classification).
Yan’s real-world example was of a publicly available aircraft engine run-to-failure data set, which showed how the models can be conveniently trained and compared with different algorithms in Azure ML. This was very interesting to me because, even though the bottom line of all this is to make money, an added benefit is obviously to improve airline safety and ultimately save lives (like mine, for instance).
Overall I enjoyed the day immensely. Not being a programmer, some of it made my head spin a little, but I definitely get the feeling that Machine Learning and more specifically, Predictive APIs and Apps are a big deal – and getting bigger all the time – and that Predictive apps will be the next big thing in app development. Don’t believe me? I’ll leave you with these thoughts:
“In the next 20 years, machine learning will have more impact than mobile has.” — Vinod Khosla, Founder of Khosla Ventures
“Predictive is the ‘killer app’ for Big Data” — Waqar Hasan, InsightsOne CEO
“If we can get usable, flexible, dependable machine learning software into the hands of domain experts, benefits to society are bound to follow.” —Dr Kiri L. Wagstaff, Researcher at NASA
Interested in reading more Machine Learning material? It’s one of Cloud Academy’s favorite blog topics!