Is machine learning real, or overhyped?
Machine Learning, coupled with the ever elusively defined field of “Artificial Intelligence” promises to revolutionise the way businesses work, cars drive, factories plan — pretty much every area of our life today. It’s a promise that the technology behemoths — Google, Amazon, Apple and Microsoft — have poured millions of dollars of research into. We know that machine learning is at the centre of key technologies such as voice recognition (Alexa, Siri), face recognition (Face ID) and autonomous driving (Tesla) — but where’s the rest of the technology application? Can we point to real world machine learning applications outside the big tech behemoths? And if not, why not?
One interesting application of machine learning is in predictive scoring — whereby an algorithm can begin to make predictions regarding an outcome. Salesforce Cloud’s Einstein product is a good example of this. Their algorithms claim the ability to predict success factors on Leads held in Salesforce Cloud. The idea here is very attractive — Einstein should be able to predict which leads are likely to close allowing Sales Managers to focus on the top 10% of their opportunities and nurture those potential clients through to closing. It should signal the end to the ‘shotgun’ approach to lead acquisition and management. Initial reviews when the tech launched back in 2016 were mixed, with some claiming the hype outstripped the reality of its usefulness. Fast forward to 2017 and the technology has picked up, with an example case study from Silverline quoting 30% higher close rates since going live with Einstein.
One of the difficulties with truly assessing the success of AI tech, however, is really understanding the driving force behind that success. Let’s imagine we take a company like Silverline who decide to embark on adding AI to their sales teams processes. The first thing that machine learning demands is data: good data and lots of it. Second to rollout is instigating change: “Listen up sales team! We’re rolling out AI to your sales software. We want you to focus on the leads that have a predicted 90%+ closure rate”. Finally, this change demands focus and adjusting the way the team operates: don’t do what you used to do, follow this systemic process because the technology demands it.
The tech goes live and close rates go up 30%. A resounding success. But pausing for a moment, and thinking objectively, can we conclusively say that the machine learning suggestions drove that improvement? Who’s to say that the mere razor sharp focus on a subset of sales (those predicted at 90%+ closure rate) wasn’t actually responsible for the improvement? Would it have mattered which leads had that focus applied or could the same level of success be attributed to the mere act of focusing on some leads and not others? How did improvements in data contribute to the success? Did the introduction of a systemic process help the chaotic sales people operate more effectively?
The real crux of the issue and certainly where this author feels we are, is that the gamut of “Artificial Intelligence” technology is very much at a handholding stage. Humans can still routinely outperform machine learning algorithms in almost every single application of the technology today. Many technologists won’t admit this but the evidence is clear — how many times does Siri fumble to understand your meaning compared to day day to conversations with other humans? Does your partner ever fail to recognise your face the way Face ID sometimes does? Does your Tesla Auto Pilot drive as well as you do? Does Salesforce Einstein outperform a seasoned sales professional?
So where does that leave us? Consensus seems to be that “AI” and particularly machine learning is currently highly effective — and impressive — at very narrowly focused tasks. This is obvious to anyone who understands how a machine learning matrix actually operates. Even deep learning is about training a network to generate predictions or outcomes that operate on a constrained set of input data, solving a highly specialised task. This is why the Salesforce Einstein technology will work well for some, and terribly for others. The training input for the model depends on the ‘law of averages’ across, presumably, all of Salesforce’s customer dataset. So if you have a slightly different sales approach it’s very difficult for an algorithm to accommodate you.
Machine Learning today, therefore remains highly impressive at highly specific tasks, such as voice recognition, facial recognition and image recognition. The question therefore is — where can this technology be applied in the real world? The answer: in highly specific tasks. And the business benefit? Automating those tasks. If it’s a highly specific, but complex task then automating that task through machine learning is the way to go. We are working with customers on projects such as classifying damage to objects and training a machine learning model to categorise and price damage using the same judgement a human operator does. How useful would it be to present a machine learning algorithm with 50 random photos of an object and for that algorithm to suggest a cost to fix it? This is completely within the realms of what is possible with machine learning today.
The best times for this nascent technology are certainly ahead. We believe that as new approaches from academia and research move to real world application, coupled with the democratisation of technology through Cloud Computing infrastructure, the number of real world machine learning applications is going to increase exponentially in the coming years.
Part 3 of our Data Driven Work Management series
Part 2 from our series of posts extracted from our “Data Driven Work Management” Whitepaper.