Machine learning has come a long way since Frank Rosenblatt discovered the perceptron algorithm and Arthur Samuel developed computer checkers in the late 1950s – having moved out of academia and into widespread industry adoption.
That’s a trend Poul Petersen, the chief infrastructure officer of machine learning API company, BigML, is seeing, noting the recent explosion of APIs coming onto the market. Petersen is speaking today at the Predictive APIs conference in Sydney.
“The goal in that time [the early days] was more about designing algorithms and defining the theory. It’s easier now to say it’s A, B and C, but when you’re the visionary, it’s much harder to define,” Petersen told CIO.
“The turning point was with the International Machine Learning Society in 1980, which hosted the first machine learning workshop in Pittsburgh, Pennsylvania,” he said.
“That’s when the field became self-aware. It was no longer just a branch of statistics; it really evolved into something that people understood they could study.”
With decades of academics defining the algorithms, fine tuning them and finding ways to boost their performance in predicting or classifying, people are now more focused on scalability and applicability, he said.
This means being able to scale machine learning algorithms on huge volumes of data, which is often needed in deep learning, and finding new applications in businesses that are not necessarily the likes of IBM and Google.
Petersen said the main barrier to widespread industry adoption of machine learning is scarce technical skills. As the co-founder of a predictive API, the whole idea is to democratise machine learning so that it’s not left stuck in the hands of specialists.
“There’s theoretical machine learning, where people are very focused on idealised problems, creating new algorithms. But it’s not enough just to invent algorithms, it’s really only useful if you can do something with them and put them into practice.
“From a business perspective, you don’t really care about machine learning, you just need the results.
"And it’s just one tiny piece in this huge puzzle. The tool is eventually going to become another part of the layer, so there’s more focus on the business problem and less on the tool.
“The focus will eventually move away from the learning layer … where everybody will use machine learning, as it’s just another part of the [computing] stack,” he said.
Exportability and freedom in moving the machine learning model script out of the API and into a local app or system is another trend Petersen believes other APIs will move to.
“The idea is that you own the model that you built with your data. It’s your data and it’s your model, so you should extract it, take it out and install it into your own software, or do whatever you want with it,” he said.
“Also, people are used to things being free. Almost everything is free - you can get a Dropbox account for free, a Google account for free. So I definitely think that any of the players now that have a machine learning API that aren’t offering a free tier, they all eventually will have to because there’s just going to be complete market resistance to that,” he added.
However, a machine learning API won’t replace the need for a data scientist or the like, Petersen said.
Algorithms cannot define a business problem, source and extensively clean up data, and look for opportunities to use data to improve the business or better solve problems.
“The difficult part is knowing what you want to do and how to get the data to do it. The algorithm has no idea, it only processes what data you give it.
“And there’s the feature selection and feature engineering. That’s a huge problem that no one has really solved in an automated way.
“There are lots of things you can do with machine learning, but who sits down and thinks about the data they have and how they can optimise their business based on their data? Probably nobody in most companies, so that’s a perfect role to have someone dedicated to."