The Real Reason behind all the Craze for Deep Learning

A simple English explanation, minus the math, stats & code

You can now read this article in Japanese (thanks to Koki Yoshimoto).

Deep learning has created a perfect dichotomy.

On the one hand, we have data science practitioners raving about it. We have the aficionados jumping in to learn and make a career out of this supposedly game-changing technology in analytics.

And then there is everyone else wondering what the buzz is all about. With a long list of cool technologies already projected as the panacea to business’ problems, one wonders what this additional ‘cool thing’ is all about.

For people on the business side of things, there are no easy avenues to get a simple and intuitive understanding. A Google search gets one entangled in the deep layers of neural networks or gets them bowled over by the math symbols. Online courses on the subject haunt one with a bevy of stats terms.

One eventually gives in and ends up taking all of the hype at face value. Here’s an attempt to demystify and democratize the understanding of deep learning (DL), in simple English and in under 5 minutes. I promise not to show you the cliched pictures of human brains, or a spider web of networks :-)

So, just what is Deep learning?

Let’s start with the basic premise of machine learning (ML). The attempt is to teach machines on how to get to the desired outcome when presented with some input. Say, when shown the past 6 month’s stock prices, predict tomorrow’s value. Or, when presented with a face, identify the person.

The machine learns how to do things like this, obviating the need for laborious instructions every time.

Deep learning is just a disciple (or, discipline) of machine learning, but with a higher IQ. It does the same thing as above, but in a much smarter way.

And, how is it different from machine learning?

Let me explain this by using a simple example of face detection.

(Pic: “Jimmy answering questions” by Beatrice Murch derivative work: Sylenius, licensed under CC BY 2.0)

Traditional face recognition using machine learning involves first manually identifying noticeable features on a human face (such as eyes, eyebrows, chin). Then, a machine is trained to associate every known face with these specific features. Now show it a new face, and the machine extracts these preset features and does a comparison to get the best match. This works moderately well.

(Pic: Machine learning is fun.. by Adam Geitgey)

Now, how does deep learning solve the same problem? The process is nearly the same. But remember, this student is smarter. So, instead of spoon-feeding standard facial features, you let the model creatively figure out what to notice. It may decide that the most striking feature in human faces is the curvature on the left cheek, or how flat a forehead is. Or, perhaps something even subtler.

(Pic: Facial features identified by few interim layers of DeepFace DL model)

The DL model silently figures out this connection between the input (face) and output (name), when shown tons of such pairs. Then, when presented with a new face, voila it gets it right magically. Compared to earlier recognition techniques, DL hits the ball way out of the park, in both accuracy and speed.

(Icons - hunotika, MGalloway(WMF), Google [CC BY 3.0] via Wikimedia Commons)

But, why do they always show pictures of the human brain?

To be fair, there is a connection.

Let's review how a child learns her first lessons. You show flashcards with the picture of an elephant, and read it out aloud. After a few such instances, when the baby looks at any semblance of an elephant, she identifies it instantly. Irrespective of the pose, color or context. We didn’t teach her about the trunk, tusk or shape of ears, but she learned it in totality. And she just gets it.

Just as we are unsure how the baby learned to identify what makes up an elephant, we really don’t know how neural networks, the technology behind deep learning figures this out. This is where all similarities to the human brain and neural connections spring up, but I’ll stop here and save you the hassle.

It suffices to know that deep learning is insanely smart at automatically identifying the most distinguishing signals (features) in any given data(face). In other words, it is a master at feature extraction. When given tons of input-output pairs, it identifies what to learn and how to learn it.

Deep learning figures out the strongest pattern in any presented entity — a face, voice or even a table of numbers.

Is this such a big deal for Machine learning?

Yes, it’s Huge.

In spite of the stellar advances in machine learning, the biggest challenge facing the discipline has been… you guessed it right, feature extraction. Data scientists spend sleepless nights discovering connections between an input (a hundred factors of customer behavior) and output (customer churn). Then the machines conveniently learn from them.

So, the difference between top accuracy and poor results is the identification of best features. Now, thanks to deep learning, if machines can do this heavy lifting as well automatically, won’t it be neat?

What use does a pattern identification machine have for business?

Plenty.

Deep learning can be applied anywhere there is a fitment for machine learning. It can comfortably thrash problems with structured data, an area where traditional algorithms reign supreme. Based on what we’ve seen, it can crash the learning cycles, and push accuracy to dizzying levels.

But the biggest bang for the buck is in those areas where ML is still stuttering without a brisk start. Take the case of images, video, audio or deeper meaning from plain old text. Deep learning has crushed problems with such data types that need machines to identify, classify or predict. Let's look at a few:

• Advanced face recognition technology is seeing early applications in the real world, and the quality of image or exposure are no longer constraints.
• It has made not just detection of animal species possible but lets us name every whale shark in the ocean. Say hello to Willy, the humpback whale!
• Advances in speech recognition cut error rates by over 30% since DL took over. And about 2 years ago, they beat humans in this space.
• DL has endowed machines with artistic abilities, and there are interesting applications of image synthesis and style transfer made possible.
• Thanks to DL it is possible to extract deeper meaning from text, and there are initial attempts to solve the rankling challenge of fake news.

(Object detection using Deep learning on Tensor flow, by Diego Cavalca)

That’s all too smooth, isn’t there a catch?

Well, the biggest advantage of deep learning is also its shortcoming. The very fact that humans don’t have to identify distinguishing features means that the machine defines what it deems important. We, humans, are creatures of reason. We have trouble with anything that doesn’t fit a mold.

Trouble brews in this paradise when one tries interpreting the meaning of machine identified features or attempts to transparently explain why a machine’s decision must be implemented. After all, how comfortable is a business decision-maker to bet millions, or worse, place lives of people at the altar of cryptic, but accurate recommendations by a tool, that was invented a few years ago?

Interpretability of deep learning algorithms and visual explanation of results is a rapidly evolving field, and research is fast catching up. And yes, it needs tons of data to even get started. So yes, there are some hiccups in this area, but the stellar and stable results clearly outweigh the cons, for now.

So, that’s deep learning in a nutshell. Please let me know what you think.

PS: This is a repost of an article from my blog on Medium.

Too much to read? Here’s a 5-minute video post of this article.