The Seven-Year AI-tch – Training Data And Defining Intelligence

“To understand how machine learning works, we only have to look at the way humans learn” explains Ganesh Subramanian, [who remains the Founder & CEO of real-time fashion analytics platform Stylumia today, as he was in 2017].  “Point to a chair and ask a young child to identify it, and unless they have previously been told what it is, they will not instinctively know its name.  If a parent or guardian reinforces the label the next few times the child sees a chair, the child will very quickly pick up that this object remains a chair every time he or she sees it, and that when they see similar-looking objects, they are also likely to be chairs.”

“Machine learning is very similar, because the machine does not inherently know what a dress, a top, or a pair of jeans looks like – we have to keep throwing example images at the computer vision portion to train it.  We steadily reinforce that this is what a dress looks like, for instance, and then the program will later be able to infer that other styles and shapes like it are also probably dresses.  The major difference is that machines find things easy that humans do not, and vice versa.  Ask a person to add up twenty different, twelve-digit numbers, for instance, and you can expect to wait a while for an answer.  A machine can perform that calculation in nanoseconds.  But point to a red curtain and ask a person what that object is and what colour it is, and you will get a quick answer; a computer that hasn’t been properly trained will not be able to respond at all.”

While the rest of this feature series is given over to the applications and impact of AI on the retail, footwear, and apparel industry, the differences in the inherent capabilities of humans and machines raise an issue that no real examination of artificial intelligence can overlook: is there something that makes us special?  More specifically, for our purposes, is there a natural aptitude for context and creativity that makes humans capable of recognising and responding to art or style.  And if so, can that, too, be taught to a machine?

To briefly address this thorny philosophical line of thinking, we’re going to revert to Google Photos as a case study.  In the interests of building an image recognition engine that can spot a mountain in a photograph, Google’s engineers will have focused on only what counts: the visual elements of a mountain.  Given access to Google’s vast repositories of data, the machine may also be able to call up associated datapoints such as the temperature at the mountain’s peak, or list the mineral composition of its rocks.  What it will not – and some will say cannot – have access to are the feelings that a photograph of a mountain can conjure up in the mind of a human being: freedom, exploration, openness, vertigo.  All of these contain an element of what philosophers call ‘qualia’ – discrete instances of subjective experience that, arguably, serve as evidence that humans, and only humans, are conscious and self-aware.

As an example, were I to look at a photograph of a mountain, I could imagine myself on its slopes, free from responsibility.  I could, also, picture myself at its peak, gazing at the vista spread out below, and considering my life choices.  I could pretend I was on the edge of that mountain, worried about my mortality if I should fall.

These thoughts and experiences all come naturally to you and I, and almost everyone would agree that they are what separates people from machines.  And while not everyone will know the philosophical term for their stance, most people would equally argue that a machine could not have that kind of conscious experience and understanding, no matter how good it becomes at performing outwardly visible tasks like recognising shapes.

Entire careers have been dedicated to tackling this issue, so I cannot hope to address it here.  I did, however, reach out to Susan Schneider, an author, TED Talk alumnus, and [at the time, in mid-2017] an Associate Professor in the Department of Philosophy at the University of Connecticut, for her expert perspective.

“For me the mark of the mental is conscious experience, and I care an awful lot about whether an artificial system can be said to have an actual experience”, Schneider told me.  “From that point of view, I think we can eventually have a super smart AI that outthinks us and outperforms us, but without it being what I’d define as conscious.  There are reasons why I don’t think information processing and consciousness necessarily go hand in hand.  For example, consciousness in humans is recognised through our ability to perform novel learning tasks, reasoning and so on – things that an artificial intelligence could theoretically replicate.  So building a machine that can also reason would get us to the point where that AI would technically qualify as intelligent, without the need to have the same kind of conscious experience that we have.  I think in the long run it’s likely to be an empirical matter: what kind of architecture a machine has will determine whether or not it’s conscious.  Our biology has given rise to consciousness and information processing, but other substrates may not be able to host consciousness even though they’re as good as we are at information processing – or better.”

This does not mean, though, that there is a hard limit on the capabilities of machine learning, or that AI programs are forever constrained to following rules – with no chance to translate their unique way of learning into some kind of creativity – as Schneider explained:

“I think AI is already more than up to the tasks of visual recognition, logic, and even extremely complex games, where an AI has already exhibited the ability to think outside the box to some extent.  The team behind AlphaGo, [a then-recent milestone in A.I. research, where a machine beat a grandmaster at the ancient, incredibly complex game of Go,] specifically chose the game because it’s extremely combinatorially difficult and is therefore a great proving ground for whether an A.I. can break its boundaries, so to speak.  And if you look at the transcripts of the people who are analysing the A.I.’s moves, they do describe some of them as creative and intuitive.  And that’s mind-blowing even to the people who developed the program.”

“One move was even referred to as the “bathroom move” because the opponent disappeared into the bathroom for fifteen minutes because of how perplexing it was.  The same goes for IBM’s Watson AI playing Jeopardy; in a way you can say that the rules of that game are known, but what the AI did with natural language processing involved making what I would call novel judgments.  And as AI gets more sophisticated and goes from being intelligent in only specific domains – like Go or Jeopardy – to being intelligent in a variety of ways, we’re going to potentially see genuine artificial creativity emerge, with or without consciousness to go along with it.”

The key word in Schneider’s second answer above is “specific”.  Today, there is an accepted delineation between two kinds of AI and intelligence research: general AI, which is an attempt to either replicate total human intelligence or create an entirely new kind of supreme machine intelligence, and specific or “narrow” AI, which is also referred to as Artificial Narrow Intelligence, or ANI.

As you may have guessed from the autonomous vehicles, chatbots, and visual recognition examples I referred to earlier, we already live in a world of narrow AI.  Today, algorithmic traders deal in more than half of all equity shares traded in the US, while game-playing ANI programs hold the world champion titles in chess, checkers, Scrabble, Backgammon and more.

Everything else from email spam filters to machine translation services is yet another manifestation of ANI – and more and more products and services are being optimised or revitalised through the use of narrow AI by the day.  As grandmaster Gary Kasparov, who famously lost his chess rematch against an AI, puts it in his book Deep Thinking: Where Machine Intelligence Ends and Human Creativity Begins, “Artificial Intelligence is on a path to transforming every part of our lives in a way not seen since the creation of the Internet, perhaps even since we harnessed electricity.”

A key point to understand here is that narrow AI is not to be considered a poor version of general intelligence.  It is, rather, a way of translating the fundamental principles of machine intelligence into manageable use cases.  And after all, in an industry like fashion, where craftsmanship and experience count for a great deal, since when was specialisation considered a bad thing?

“All the current applications of AI in our industry are narrow,” I was told by Courtney Connell, [who was then Marketing Manager for lingerie brand and retailer Cosabella, and who has since left the fashion industry]. Cosabella [in 2017] used an AI platform to completely redefine the way it markets and sells its products online.  “I wrote a piece for Womenswear Daily about AI in general, and afterwards someone contacted me to say that I was wrong to be discussing limitations, because an AI had recently beaten a professional poker player by bluffing, so that meant the sky was the limit.  The person who contacted me wasn’t technically incorrect, but he was repeating a widely-held misunderstanding about the value of specialisation.  If an AI can read facial micro-movements and tone of voice more accurately than any human being can, then of course it’s going to beat people at poker.  The problem is that people look at these narrow applications and conclude that results from a single hyper-focused program mean that every AI everywhere is now able to achieve the same results.  It’s important to remember that commercial AI is a tool, built for a particular purpose, and is not necessarily all that useful outside its particular focus.”

While specialisation is valuable in its own right, where extensive domain expertise counts the most is in specialist markets.  In these industries – and fashion is unquestionably one – the ability to understand minor variations between thousands or millions of different products is essential.

“Take the automotive industry,” explained Eric Brassard, a former executive at Saks Fifth Avenue, and [in 2017] CEO of Propulse Analytics, which [was] using deep learning to reinvent product recommendations.  “When it first started, 120 years or so ago, any vehicle with four wheels and steering wheel was a car.  Today there’s a significant difference between an SUV, a Ferrari, and a regular sedan in most people’s eyes.  In fact, there’s an equally significant difference between a sports sedan and a comfort sedan.  As any industry evolves, specialism breeds complexity and subjectivity, and making sense of those, at scale, requires the appropriate, specialised solution. We started from the belief that AI is the right solution to the problems of complexity and subjectivity in highly specific industries, and our position has not changed since.  Right now no AI is omnipotent, and the ones that are highly potent are only potent in their highly specialised fields.”

Specialisation in AI also has compounded benefits.  As an algorithm becomes better able to spot micro-scale variations in fit, component placement, hem length, grading rules and so on, it gains the ability to plot these changes as trends over time, and to then refer back to its own research and conclusions in order to draw further conclusions.  “True AI systems are self-learning, which goes a step beyond machine learning,” explained Andy Narayanan [who was then VP of Intelligent Commerce at Silicon Valley AI company Sentient, and is now, a propos, Head of Product for Generative AI and Foundation Models at Apple].  “Looking at computer vision exclusively as a way of using an algorithm to conclude that this is a blouse, that’s a belt, and so on, is limiting its potential.  Being self-learning means that the algorithms should be capable of adding another dimension, taking a view of how blouses and belts are changing over time, and determining what kinds of each are likely to be in fashion today – and why.  This is where I think the true promise of AI is: understanding what features, facets, colours and textures a product category has now that it didn’t have before, and adapting its recommendations to that evolution in real-time, with no human intervention.”

Another corollary benefit of specialisation, of course, is trustworthiness.  Generally speaking, we tend to put our confidence in people who we believe know what they are talking about.  This is, of course, also particularly true in specialised industries, where a new entrant has little choice but to rely on the experience and expertise of long-serving specialists.  And just as we have faith in human beings who exhibit that kind of specialisation, the same is already proving to be the case in other industries like financial technology (or “fintech”) where account holders in our native UK are trusting consumer-facing AI platforms to analyse and advise on their monthly spending.

The most interesting of these platforms is Cleo (currently available in the UK only at www.meetcleo.com), which marries a friendly conversational interface with deep learning that identifies spending patterns, trend lines, savings opportunities and other insights that the account holder would be either unlikely to research on their own, or would perhaps not even be capable of analysing.  I road-tested Cleo as part of the research that went into this publication, and that test revealed two insights that surprised even me: first, I am willing to accept a lot more intrusion from an AI than I would a human; second, I trust what Cleo says implicitly.  The application may not have the best user experience in the world, but when it shows me a trend line of my savings over the last quarter, my instinct is not to argue, like I might with a person providing the same information, but to accept that her (for Cleo is a she) specialist knowledge trumps my own arms-length acquaintance with my own income and expenditure.

Drawing these kinds of parallels between AI and human experts does, however, raise the question of how – and how transparently and reliably – both acquire their experience and expertise.  After all, conclusions reached on the basis of inaccurate data, however intelligent the analysis, are still incorrect conclusions.

We are all familiar with the way this process works in humans – we work diligently to build a career, collecting experience as we progress, and eventually narrowing down our specialisms until our knowledge of a chosen domain is deeper than others’.  For a machine, the spirit is similar, but the speed and the method differ dramatically.  For some purposes, specialism to the requisite degree can be achieved by letting one or more algorithms loose on a large data pool, or presenting them with regularly-repeated images on a similar theme until they demonstrate the ability to distinguish their contents.

More often than not, though, the process of specialising an AI is actually a human-led one, whether the humans are aware they are participating in a learning activity, or whether it is tacked on as a by-product of another product or service. Google’s recent changes to its CAPTCHA system (the gatekeeper of comments forms and login screens everywhere) have seen users asked to identify squares of images, clicking all those that contain taxis, or churches, or street signs.  And while the search giant has not exactly admitted this, the results are all but guaranteed to be going towards the specialisation of algorithms designed to recognise these things.

But this wide-net approach is not viable in situations where the luxury of testing at scale on an audience that’s ignorant of your intentions does not exist, or where guesswork or incorrect answers on behalf of the human training pool are unacceptable.  “We provide training data as a service, helping companies developing AIs or implementing machine learning to get their data into the right shape to train those systems,” explains Daryn Nakhuda, [who was then Founder and CTO of Seattle-based Mighty AI, which prided itself on being able to educate AIs through both public, vetted, training exercises and domain-specific data collection and training for fashion, retail, and autonomous vehicles].  “Everyone talks about big data, and how much of it they have, but not a lot of it actually ends up being used.  To get around this, we identify particular attributes that a client wants to be able to search or perform recognition on, and figure out how it can be categorised and organised into some kind of taxonomy.  In a lot of cases, the data that contains those attributes is unstructured, coming from social media or other similar sources, and isn’t provided in a form that an AI can easily use.”

“To that end we also have a dedicated web application, Spare5, that allows real people to perform different micro-tasks that help train these AIs.  For fashion that might be identifying the individual garments in a photo, characterising the fit of an outfit, or distinguishing between formal and informal wear – things that businesses have never been able to model at scale before.  We’re using humans to make judgments that machines are not able to make, because, with the right people making the right kind of subjective decisions or identifications, the machines will eventually learn to make those decisions themselves, and will begin to identify patterns that no human would be able to pick out.”

As Nakhuda’s answer suggests, experience on the part of the training pool is essential.  In short, we cannot expect to create a specialist AI without specialist educators.  While the general public, for example, may be able to successfully pick out certain elements of an outfit, their ability to distinguish between different fits, different weaves, different materials, and other metrics will be limited.  And, equally importantly, their ability to correctly identify key criteria will be based on how closely they align with the retailer or brand’s customer demographic.

“We’re not necessarily looking for fashion expertise in our trainers, but rather an alignment with the retailer or designer’s target audience,” added Nakhuda. “We can certainly qualify people by asking them to recognise the difference between an Oxford collar and a spread collar, but it’s equally important that we qualify them on the basis of cultural fit.  A retailer needs to know that if their AI is going to describe an outfit as appropriate for work, it’s appropriate for work in their target market.  What we wear to the office here in Seattle, for example, is pretty different to the way people dress for work in Miami or Europe.  And the same goes for women’s clothing, where styles vary dramatically depending on climate and culture.  You don’t want your AI to be a one-size-fits-all system; you want one that will share perspectives and experience with the people who will interact with it.”

So where now for AI, specialised, trained with appropriate human knowledge, and capable of bending the rules from time to time?  A common thread among all the interviewees we solicited for these stories is that general intelligence remains a pipe dream – and that the quintessential human experience may never be transferred to a machine.  This does not, however, preclude the development of significantly advanced AIs who do not require consciousness in order to exceed our every limitation, or put a limit on the potential of the best ANI solutions to achieve mass adoption in their specialist markets – whether they are in mass market fashions, media or the military.

Indeed, outside our industry, AI is already being pegged as the next international arms race.  Just a week before this story was written, Russian leader Vladimir Putin, asked about the future of AI went on record as saying that “the one who becomes the leader in this sphere will be the ruler of the world”.  As endorsements come, this is perhaps an uncomfortable one, but the message is clear: after several generations of being confined to stories and research labs, AI can no longer be ignored.

It is unlikely, though, that you, the reader, has world domination in mind.  More likely, you are wondering just how the dawn of the intelligence era will impact your business.  And while this primer has been far-ranging, let me assure you that everything contained in these pages should be considered grounding for understanding a technology that I guarantee will transform your day to day life at work and at home far sooner than you may expect.

“At this point in time general AI is very far out, and I honestly think it still resides in the world of fiction,” added Andy Narayanan [then of Sentient, now of Apple].  “Mass extinction of jobs because of the creation of a super intelligence just isn’t likely to happen any time soon – and any general AI might prove to be too broad to actually solve immediate problems.  A better application of AI is instead to solve a very thin slice of problems at a much deeper level – that’s where the applications are today.  There are a variety of areas in the apparel business alone where human decision making is sub-optimal, which we are now handing off to AI, because we know that AI can do them better and faster, and can do them at scale.  Instead, humans are going to work on something more creative.  Personally, I feel like this is the best way to make AI successful.  Rather than having these high expectations of a general or super intelligence that we might not be able to meet in even the long term, we’re better off setting narrow goals that have a big business impact.  The best thing we can do is use AI to augment human capacity in ways that deliver hundreds of millions of dollars of value for retailers and brands here and now.”