The beginners guide to machine learning
“The last 10 years have been about building a world that is mobile-first. In the next 10 years, we will shift to a world that is artificial intelligence (AI) first.” - Sundar Pichai, Chief Executive Officer, Alphabet (Google)
There is so much discussion about AI, or machine learning, that the general population would be excused for going a little glazed in the eyes at its mention.
People can touch and feel a mobile phone, so understand its importance. Mobility has become part of the infrastructure to support high bandwidth services – video, music, search engines, all of which helped create the diversity of the internet and make it a good place to spend time, buy things or build a business.
But what of the data that is being accessed with the mobile device? How should we think about that? The answer is that every time someone is tagged in photo, or uses Siri to make a phone call, or uses a digital map to navigate, AI (a term which can be used almost interchangeably with machine learning) is part of the process.
Connected devices everywhere are the prelude to the rollout of the machine learning services which will change the way people live their lives – which is what Pichai saying.
And for better or worse, it is business that is driving the adoption of machine learning.
In fact, it is so widespread that it is a key part of virtually every significant datacentre and cloud business in the world – Amazon, Google, Microsoft, Alibaba, Tencent, Baidu, Facebook to (name just the big ones).
For example, many Americans already know Alexa, the voice activated speaker which makes soap and nappies arrive with the vodka order on a Tuesday. But they mostly don’t know that Alexa moonlights as Lex, and he (it) provides the processing power to run the annoying chatbot which keeps you busy ahead of handing you on to someone that can actually answer your questions.
And how much does this cost? Below see the Amazon Web Services pricing structure for the Lex product:
So that’s 4000 speech requests and 1000 text requests for US$16.75.
How about facial recognition software? Amazon offers this for free for the first 5000 images per month, for one year. After that, its $1 per 1000 images processed, or $1000 per one million images. In some cases, Amazon services are even billed by the second. Machine learning is so widespread that there is pricing matrix for it, and getting a taste of the services costs no more than a six-pack.
Who might use machine learning? Well, any retailer wishing to greet the customers in the store by their first names (assuming, we hope, opt-in). And who might use speech recognition? Any customer-facing business trying to improve its, well, facing of the customer.
How about using machine learning in fraud protection? An executive of a company involved in this is quoted on Amazon’s site saying “In order to counter evolving forms of fraud, we needed to build and train a larger number of more targeted and more precise machine-learning models. Once you start catching a form of fraud, the fraudsters themselves will change their strategy—so it’s a constantly evolving problem.”
Meaning that if you may make an acquisition in Brisbane, and a few seconds later log a different transaction, but for the same amount, which bobs up in Ireland, it can be hard to tell whether it is fraudulent. Sure, you and your Uber driver will know, but multiply it by the millions of transactions per second, and you can see where this is heading.
Malcolm Turnbull talks about cross-referencing the feed from the nation’s security cameras with the licence photos of its drivers – a machine learning problem requiring a massive dataset of images integrated in real time with video capture, for deployment at times when crowd safety is an issue.
It’s a fact that the companies with the largest datasets (ie customer information) have the ability to generate the best information on those customers, as long as the tools are there to support it, which is the game the big disruptors are playing. Data has already been called the new oil by The Economist, as we wrote here.
But it’s also true that making AI available on demand is good if you are a small business, because it leverages your otherwise limited resources to provide transactions.
Naturally, there are some downsides. Snapchat, which is one of Google’s largest datacentre customers, is paying US$400m a year for the use of that processing power – an amount that the company just lost in the quarter. At its worst, the machine learning datacentres have the ability to effectively levy a tax on every user of their services, which is shaping up to be a lot of users. Price-gouging is a very real possibility, as is cartel behaviour.
Investment in data companies is a key part of the holdings of Loftus Peak. We think it has a long way to run.
For further insights from Loftus Peak, please visit our website