The Seven-Year AI-tch – Our Analysis Revisited

Ask someone to name a science fiction concept, and the odds are good that they will come back with either domestic and industrial robots that are indistinguishable from humans, or a benevolent (or malevolent) supercomputer that comes to rule the world.  These themes have recurred in future-set tales for more than a hundred years. They are popular because, as with all good speculative storytelling, they ask us to consider what our place, as thinking, breathing humans, is likely to be in a world where either our physical or mental capacities have been outmatched by machines.

The upshot of having such a body of sci-fi dedicated to these subjects is that some of the smartest people in the world have spent a great deal of time considering their implications.  Some putting pen to paper; others toiling away in laboratories and research centres to bring the ideas to life; each spurring the other on with fresh ideas.

Given how well the workings of our bodies are understood, compared to the almost complete opacity that shrouds the way we think, you might expect science to have made the most progress on those physical robots.  In practice, the opposite is true.

There is a name for this subversion of expectations: Moravec’s Paradox.  Outlined by roboticist and futurist Hans Moravec in the 1980s, the paradox is expressed like this: “it is comparatively easy to make computers exhibit adult-level performance on intelligence tests or playing checkers, and difficult or impossible to give them the skills of a one-year-old when it comes to perception and mobility.”

In brief, what Moravec meant is that our mental processes and actions – the results of a consciousness, individualism, and autonomism that generations of scientists and faith scholars have been unable to adequately define – have proved far easier to replicate than the coordinated motions of our bones, tendons and muscles, which entry-level biology has long since explained away.

The mobility side of Moravec’s paradox is still a challenge.  Even the best contemporary attempts at having robots mimic human face and body movements come off as unsettling, and companies have, by and large, paused – if not abandoned – their attempts to make robots that look and move the way we do.  While people might make good models for thinking and problem-solving, it transpires that there are far more practical examples to follow for picking things up and moving them, navigating through warzones, or conducting fine-grain assembly tasks.

We have, however, made considerable progress on having machines pass common (and even complex) intelligence tests, as well as on introducing techniques that allow for computers to perceive and recognise the world around them.  These advances all fall under the broader umbrella of Artificial Intelligence (AI).  And that’s an umbrella you have more than likely seen pop up in the news a lot lately.

“Japanese company replaces office workers with artificial intelligence,” reads a typical AI headline from The Guardian at the beginning of 2017.  In isolation, it may sound alarmist or unbelievable, but it is accurate: Fukoku Mutual Life Insurance in Japan made 30 experienced staff redundant because their jobs, calculating policy payouts, could be done cheaper and better by an A.I.  Fukoku expects this to increase productivity by 30%, and to become a profitable investment in less than two years.

Maybe it’s because automation and the resulting mass layoffs are happening on a much grander and more pressing scale in manufacturing, but I find it remarkable how unremarkable these kinds of stories have become.  They hit the headlines and fade quickly, inspiring fear in other insurance workers, perhaps, but not registering on the tectonic scale you might expect given that AI programs are now putting even white collars out of work.  To put this into context, Moravec’s Paradox talks about the likelihood of replicating the intelligence of a one-year-old; I have a baby in the house as I write this, and I’m not expecting him to be performing any insurance calculations or underwriting policies for a while yet.  Seems as though we’ve moved rather rapidly from crawl to walk to run there, doesn’t it?

Perhaps that aforementioned wealth of sci-fi literature has made these kinds of things feel mundane. Or perhaps it’s because we all already interact with AI of one form or another almost every day – whether we realise it or not.

Self-driving cars are, of course, the most obvious example.  Hugely complex arrays of hardware sensors and interpretation and analysis algorithms, autonomous vehicles are a sci-fi trope writ large, and will soon be hitting streets near you if they haven’t already.  Less obvious but no less valid as applications of AI are chatbots (that initial gateway of interaction you sometimes encounter on a company’s support or sales channels before reaching a human advisor) and visual recognition platforms like Google Photos, which uses algorithms to tag your photographs with their locations, subjects, settings, and a host of other information so you can later search for a sunset or a smile.

Each of these examples, however, has something in common.  All, to my mind and the minds of many actual experts, fall short of the definition of intelligence.  In fact, these and other examples like them are only really categorised as AI because, for the 99% of us who are not intelligence programmers or philosophers of mind, a catchy acronym is needed to capture the potential of a raft of different technologies.

“When I’m talking to a business audience, I use AI as an umbrella term, and under that umbrella are specific applications of AI like cognitive services, machine learning, deep learning, neural networks and so on,” said ShiSh Shridhar, [who was at the time of this interview a Director of Business Development, Data, Analytics and IoT for Microsoft, and who now leads Retail at Microsoft For Startups].  “For a business audience, that’s a confusing set of terms, and to complicate things even further, a lot of people use them interchangeably.  As a result, I stick to the top-level term rather than having a brand or retailer worrying about the small nuances that separate individual components like neural networks and deep learning.  I realise it’s a big umbrella, but it’s an appropriate one.”

While ShiSh’s advice is sound, we are, however, going to have to talk about these components individually at least to begin with, as foundations for the visible outcomes of AI. It is only by understanding the essential building blocks of machine intelligence that we can assess where and why it exceeds, equals, or falls short of human intelligence – and why, in many cases, a machine not being truly smart does not necessarily matter in the open market.

“I think machine learning is the preferable term for what we’re talking about today,” said Julian McAuley [who was an Assistant Professor at University College San Diego in 2017, and who was elevated to a full Professor there in 2021], whose research into behavioural models and visual recognition caught my attention as I was gathering material for these features.  “The research I do would probably be called AI by the public, and even by our department here at the University, but I think that term is overused.  We are not specifically trying to build up a general intelligence; we’re taking a data-driven approach to solving a specific problem – nothing more.  When people shop online and are shown product recommendations or adverts that they feel are uncanny, they tend to assume that an AI recommendation system is truly intelligent and knows something intimate about them.  The truth is less exciting: all the typical algorithm needs to know to do its job is that one person performed an action that was similar to another person’s action, like buying a product, and it can recommend them both similar other products as a consequence.”

McAuley’s example, as well as the others I referred to before – photo categorisation, self-driving vehicles, and chatbots – are good examples of the state of commercial AI today.  They seem intelligent at first glance, but their smarts are actually confined to very small, very specific areas.  To choose the right pre-canned responses, a chatbot need only concern itself with the customer’s immediate concerns.  To recognise a mountain range in a photograph, Google Photos will look for indicators that a mountain is present, but it cannot be said to actually know what a mountain is, and neither does it need this additional context to achieve its purpose. Generally intelligent these things are not.

Let us shed a little more light on why this is the case, and why the distinction matters.  To recognise that mountain, Google Photos (or an equivalent, although Google leads the pack in this area) uses neural networks, which are algorithms arranged in an approximation of the way we believe our brains work, designed to allow machines to look at things, and then to remember and learn from them.  These are arranged in a way that permits what is called deep learning, which means that the results the end user receives are arrived at from the separate or sequential input of several – or hundreds – of layers of hyper-focused neural networks.

One such network layer might detect naturally-occurring edges and nothing else; one might look for the texture of a rock face; another might scan the top portion of every image for sky, and the portions below that for peaks and snow.  At the top level, yet another neural network will take the composite output of the nets below it and conclude, with 90% certainty, that there is a mountain in the photograph at hand.  If the algorithm has a library of other mountain images to compare this one to, perhaps it will also go away and try to match its peaks and valleys to other photographs, before inferring that this picture was taken in Yosemite.

Make no mistake, this is an exciting, road-tested application of a class of technologies with massive potential to change the way we live our lives – particularly when we realise that its use need not be confined to static photographs. Recently, Microsoft unveiled a research project it calls “Seeing AI,” which uses the same principles as other photo recognition software, but applies them to both pre-recorded and real-time moving video as well as still images.  In use, a deep learning solution evaluates the real world through a smartphone camera, and turns this hugely complex data feed into synthesised voice narration designed to help blind and partially-sighted people to navigate their environment, recognise people, read mail, and discover products.

Let me be unambiguous about this: I find the Seeing AI application incredibly impressive.  I am only 35 years old, but this is one of several occasions where I have been reminded of just how much has been achieved in my lifetime.  Using a handheld computer more powerful than the mainframe that used to occupy almost the entirety of my father’s business premises, and a camera sensor the size of a fingernail, a partially-sighted person can now have an AI in the cloud on the other side of the planet describe the world to them in real-time.

But while machine learning systems like these are terrific at the tasks they are assigned – looking for recognisable things in pictures or moving video – a significant amount of work has gone into achieving this level of recognition.  As the name suggests, machine learning systems are not created with intelligence and understanding built in, but rather acquire it through the same processes that human beings do.