A new era of machine learning

March 2016

As we embark on a new era of machine intelligence, Mike Lynch discusses the reality of unsupervised machine learning and its applications to the world around us.

The human brain is the most complex structure yet known in the universe and our perhaps our key evolutionary advantage over all living things. With an average mass of three pounds and a tofu-like consistency, the brain is a powerhouse consisting of over 100 billion neurons transmitting signals across some 1,000 trillion synapses at fractions of a second, processing everything we experience from cognition, emotion, language and memory.

Receiving vast amounts of sensorial data every moment, from the smell of smoke to the visual perception of red, the brain computes our environment and governs our response. In the tech world, this is comparable to processing reams of code distinctive only by one character per line. The sheer quantity of information we receive at any given time is colossal and would incapacitate any human who were able to pay attention to it all.

Thus, what is just as astounding as the brain's ability to merely register the richness of information surrounding us, is its ability to extract from it effectively. As humans, we need to understand our environment in a meaningful way; knowing how many leaves on a particular apple tree is useless when all we want is an apple. But how we – and by ‘we' I mean ‘our brains' – enable us to do this? It all starts at the start.

Even before we breathe our first breath, humans innately learn patterns, laying the very foundations of our knowledge. With limited visual capacities we make inferences by exploring the world through any means possible; mainly through taste, smell and touch. Each perception we have fits together in a certain way forming a manifest relational framework. For example – and I will be glossing over generations of metaphysical debate here – when we see a horse (without yet ‘knowing' the concept {horse}), we register its size i.e. bigger than a golf ball, smaller than a house, has 4 legs, a tail and is covered in hair. Notably, there are many varieties of horse, with various shapes and sizes. In addition, our perspectives of horses vary; we may see one in the dark, far away or smell one when it is wet. But through all of this changeability a number of things will always stay constant over time. We experience these things coming together in the same way time and time again until we infer patterns between them to create concepts i.e. {horse}. This pattern recognition process is how we learn and extract meaning from our experiences and develop our individual behavioural repertoires and personalities.

‘So what does all this have to do with machine inteligence?'. Well, just imagine if a machine could do all this. It is extraordinarily complicated, and requires incredibly fast processing power, and yet, we are seeing nascent examples of this every day: recognising fraudulent payments in real time or targeted marketing campaigns from advertisers are a basic version of artificial intelligence.

A key difference, however, between the brain's processing and the machine's is what we call ‘supervised self-learning'. In the case of the human brain, it is trained without input from anyone or anything else or any prior knowledge (philosophical controversies aside). In comparison, much of the machine learning in existence today is supervised, requiring pre-programmed rules in order to work.

However, we are at the dawn of a new era, where the promise of unsupervised machine learning, for so long a chimera, is becoming a reality. Much has been written about the possibilities of robots taking over the world, or of machines becoming so sophisticated that they go out of control. In fact, I think that these are extraordinarily unlikely outcomes, certainly in my lifetime. However, what we will see very soon, is pressure on middle class jobs which will be done by intelligent machines instead. Technology will be deployed to analyse thousands of documents in electronic data rooms, throwing up the riskier contracts much faster than a clutch of paralegals thumbing through them. Computers make much more accurate predictors of fraud than humans, as they don't require hypotheses to work, they learn the behaviours without assumptions.

Machines are encroaching on white collar jobs faster than you can power up your laptop. They don't get tired, they don't make mistakes, they don't take days off but more importantly, they get cleverer over time. We will see them in customer-facing roles, listening to phonecalls, providing answers to questions whilst learning everything about the caller to prompt richer interactions.

In technically complex areas, such as cyber-security, where the cyber-threat landscape is fast evolving and we simply cannot know what the threat looks like in order to protect assets, unsupervised self-learning presents a phenomenal opportunity to potentially neutralise unknown threats.

This is the dawn of the new area of Machine Learning. It is exciting for those of us in technology who have been working on this for decades, and it is exciting for society as we harness the power of this new intelligence and the insights that it brings. However in all this enthusiasm, we should also remember Daleks can't climb stairs.

A version of this article appeared in Cambridge News.