Defining Deep Learning, Part I: What It Is and What It Solves

What’s “Deep Learning”? It’s a class of artificial intelligence, of course. Though saying that is kind of like labeling the Earth as just another planet.

Nobody would confuse us with Saturn or Mercury except the most nearsighted – or indifferent – spacefarer. Within “artificial intelligence,” there are numerous and multiplying variations, and Deep Learning is just one of them.

To be specific, it’s a category of machine learning algorithms that operate on a neural network, which is composed of stacked layers of “artificial neurons”. Each neuron is a non-linear (“curved”) function that outputs a decision value (typically between 0 to 1.0) based only on values sent from the layer of neurons immediately below it. The lowest layer reads the input data (a set of numbers which can be from measurements, image pixels or text).

The signals are propagated ahead, layer by layer until the final output value emerges at the top: the neural network’s prediction.

The network’s ability to learn lies in the varying strengths of each neuronal connection, which are called “weights”. Each weight value expresses the importance of a lower neuron’s signal to an upper neuron. A trained network will have the optimal set of weights to compute accurate answers from new input data.

How is a network trained? When given input data with a labeled answer (meaning it’s already been classified), the deviation between the network’s predictions and the actual answer produces an error signal. The error signal and non-linearity of each neuron’s decision function tells us whether to increase or decrease each weight. The error signal gets propagated backwards all the way to the lowest layer. Over many training examples, the network weights are repeatedly tuned until finally reaching some satisfactory benchmark, such as accuracy level.

This training process is akin to the human brain’s tendency to rewire connections when learning concepts. Just like its biological counterpart, the behavior of an individual artificial neuron is simple (and easy to compute) but the network’s intelligence comes in aggregating numerous layers and training their connections to ultimately make complex decisions.

Learning to interpret difficult data

A key characteristic of modern neural networks is the depth of their many layers. Deep neural networks not only classify data or make predictions, but the weights of upper layers exhibit an inner representation of the data that it’s processing.

It’s a phenomenon we AI wonks call representation learning. Through it, raw data is transformed into features: distinct, measurable properties of what’s being observed. A machine can use those features to do specific tasks, but it can also learn features on its own.

Why does that matter?

In machine learning systems, many tasks demand input that’s convenient to process. But real-world data like images, video and others is typically inconvenient: complex, highly variable, unlabeled and filled with redundancies. It’s too messy to work with in its raw form.

Let’s say you’re trying to build a model for identifying spam, and you’ve gathered a lot of training data - data that’s used by the system to construct or discover predictive relationships. You’d pick features or variables that both represent the email, and are also predictive of whether or not it’s spam.

So representation learning uncovers the measurable features, or “representations,” in the raw data, so the system can now classify and analyze it.

Where’s a great place to find an outstanding example of how Deep Learning solves problems in the real world?

Let’s hit the casino!

How to face a recognition issue

By now, everybody knows there are scores of cameras watching our every move as we’re hitting the tables or feeding the slots at nearly any modern casino. 

For years, those surveillance systems have enabled gaming operators to use facial recognition software to identify who’s who on the floor, and spy out visitors ranging from card counters to known pickpockets to, one would think, the all-too-welcome ‘whales’ who just flew in from Osaka or Shanghai with huge bankrolls burning a hole in their pockets.

The way facial recognition originally worked in situations like this was by processing the image of a person’s face through a set of algorithmic filters. Those filters took into account hair and eye color and, especially, facial landmarks: the position, size and shape of the nose, eyes, cheekbones and jaw.

The issue, though, is the same one we described above: Getting a good match with this approach requires “convenient” data – a well-lit, full-frontal shot of a person’s face, and a database to match a “faceprint” against.

But. What if they’re grinning, or wearing a hat, or you can only capture an image of them in profile, or you simply can’t get a square-on image of the person? Since the data “features” we’re trying to measure in this approach are rigidly circumscribed, we can’t identify this person.

In Deep Learning, the machine learns and defines the data features itself, and doesn’t need to rely solely on characteristics like the spacing of facial landmarks. It can process even partial pictures. 

In the case of Google’s facial recognition system, it educated itself about how to best recognize faces by going through millions of training cycles using thousands of images. Eventually, it learned how to reliably interpret 128 key facial measurements. So even very different images of the same individual will still contain enough data for the platform to identify that person.

How a specific Deep Learning application like facial recognition gets used is opening up various ethical and legal questions, as Google is learning. But what are the odds that casinos – just like the military, security services, even Facebook – have already upgraded to the latest and greatest in facial recognition tools, powered by Deep Learning? Like they say: Never bet against the house.

Deep Learning enables actual ABM

In facial recognition, raw data has been reduced to classifiable numbers. There’s a similar modeling process at work when Deep Learning gets applied to marketing challenges like lead generation and customer engagement.

Even if they have access to large volumes of data, a marketer may not be able to extract anything useful from it. It’s raw data, and needs to be crunched thousands of times over to find interpretable features.

Only a Deep Learning platform is capable of processing that much information, using text filtering and topic modeling; the latter is a key statistical tool for identifying “topics” – relevant clusters of words – among a collection of documents and data. In our own application of Deep Learning, those collections are measured in terabytes, assembled from customer, social and third-party sources.

The carrot that Deep Learning dangles in front of CMOs and CTOs mulling the potential of account-based marketing (ABM) is this:

It lets the same modeling and profiling process work across multiple targeted accounts, rather than forcing a marketer to develop a separate model for each of them.

Again, that’s because it’s using representation learning to teach itself which salient features it should look for within each account, as well as among the people who work there.

That enables ABM scalability which is, you can rest assured, a good-sized carrot for B2B marketers.

In our next Defining Deep Learning post, we’ll drill down into exactly how Deep Learning gets applied to persona-building, lead generation and personalization for ABM. So stay tuned!