Deep Learning, Artificial Intelligence and Your Not-So-Distant Store of the Future

Arun Nair
Arun Nair
RetailNext co-founder and CTO

Deep learning and artificial intelligence is rapidly moving from the domain of science fiction into our everyday lives, and it has exciting ramifications for the manner in which humans do nearly everything, from work to play, and yes, even shopping.

Like most any other topic, when it comes to artificial intelligence (AI), media accounts tend to focus on big, bold and very sexy stories, like autonomous self-driving cars, or machines beating world champions at games like chess, Go and even television’s Jeopardy. And, let’s face it, those stories are pretty cool, and they probably deserve the attention they get.

But, what about AI and its role in retail? Currently, we have AI solutions like Amazon’s Echo, Google’s Home and Apple’s Siri making real differences in the online shopping experience, and the Amazon Go concept store appears destined to be a disruptive force in the brick-and-mortar realm. But, there are a lot more applications, particularly those coming online line “behind the scenes,” and they will have a tremendous impact in all areas of the retail enterprise.

What is Artificial Intelligence (and what is it not)?

Decades ago, early in the study of artificial intelligence, an interesting discovery was made by researchers. Coined as Moravec’s paradox, it stated that high-level reasoning requires very little computation, but low-level sensorimotor skills require a tremendous volume of computational resources. Hans Moravec wrote, “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.”

Written a little differently, a computer can easily dispatch the world’s greatest chess champion, but it can’t match a one-year old human’s – or even an animal like a domestic dog’s – perception, like the ability to recognize human faces and forms.

For example, with computer vision, computers sometimes have difficulty recognizing people wearing feature-less, untextured, single-colored clothing, like burqas. It’s very easy for a person to recognize that image as a person. But, in a way, that person is sort of cheating, because he’s using the world’s best pattern recognition engine – the human brain.

The field of study and the term artificial intelligence effectively came to be in 1956 at a conference at Dartmouth College, and from the very beginning it was more than a little bit confusing and every bit as much controversial. At a very simplistic level, practical AI came down to two schools of thought.

Originally, most researchers thought the best approach to creating AI was to write enormous programs that laid out both the rules of logical reasoning and sufficient knowledge of the situation, or the problem, to be thought about and solved. So, for example, to translate English to Japanese, a program would require the grammatical rules of English, the entirety of all the definitions of every word in the English language, all the grammatical rules of Japanese and then, of course, all the definitions of each word in the Japanese language. At that point, the program would be empowered to receive words, phrases and sentences in the source language of English and translate it into Japanese.

Sounds like a lot of work, right, and for just one application at that. Plus, there’s a big limitation: what about applications where the “exceptions to the rule” very often outnumber the rule itself, like, for instance the often-confusing English language? It’s no surprise pioneering AI work came in rules-defined domains like chess – remember Deep Blue’s series of matches with Garry Kasparov in the late 90s?

The dissenting view of AI is actually the AI community’s first view, dating from the 1940s, believe it or not, and is based on computers learning from the ground on up, from data, rather than learning from the top on down, from rules. It’s the principle on which the most intelligent machine and pattern recognition engine on Earth – the aforementioned human brain – is based. This view of AI is based on what’s called neural networks.

Neural Networks and Deep Learning

For any geek readers out there, the average brain contains approximately 100 billion neurons, and each neuron is connected to up to 10,000 other neurons, meaning the number of synapses – the pathways between neurons – is between 100 trillion and 1,000 trillion. We’re pretty far from building a computational neural network of that size, but not as far away as you might think, as organizations like Google have created artificial neural networks comparable to the brains of laboratory mice.

To simulate the human brain, programmers build complex networks of neurons, millions and millions of them, spread across multiple layers – those multiple layers are what led to the term “deep learning.” These layers, in turn, communicate and collaborate with one another to solve a problem. You know, human brain-like.

The difference is how the programming is performed. It’s not like traditional programming where an engineer tells a computer what to do. Rather, the engineer teaches a computer what to do, like teaching a child or training a dog.

Of all people, gamers essentially led the advancements of neural networks – bigger, better, faster and more cost-effective graphic processing units (GPUs) were developed for their gaming needs, and it’s that enormous computational speed and power that brought deep learning to life.

The future is now

Deep learning is ubiquitous and nearly everyone is using neural networks, including, likely, you, through applications like Google Translate and Snapchat, or with healthcare diagnoses, self-driving cars and even financial markets where your retirement savings are nestled. And, deep learning is ready for prime time too – two years ago, a Microsoft algorithm became more accurate than humans in classifying over 300,000 images into one of over 300 descriptive categories in the ImageNet challenge.

So, where are you going to see deep learning and AI pop up in your retail environment? Perhaps first, you will see tremendous increases in accuracy in shopper demographic data, particularly as computers are trained as to what different ages look like, particularly in the diversity-rich regions around the globe.

Still others will benefit from activity detection, understanding exactly what activities are being undertaken inside the store, by whom and how they’re being undertaken. It’s not only about understanding the differences between actions of shoppers and employees, but also what those actions are – for example, crouching, bending, reaching overhead and the like – and what aids are being used – carts, bags, brooms, mops, etc. Over time and at scale, that data delivers tangible insights, leading to strategies that deliver differentiated shopping experiences, and that’s important to us all.

Other use cases for AI in retail include better predictive models, essentially answering questions like what will happen in store tomorrow or this weekend, and automatic recommendations, along the likes of where new stores should be opened, what the layout should entail, and what products should be stocked in the store and where.

Deep learning and artificial intelligence is rapidly moving from the domain of science fiction into our everyday lives, and it has exciting ramifications for the manner in which humans do nearly everything, from work to play. Savvy, winning retailers will recognize the technology solutions are much more of a today thing than a tomorrow thing.

Join the #retail, #inspiringretail & #SmartStore conversations on Twitter @RetailNext, as well as at