The Deep Learning Age Of Computing

How Deep Learning Will Fuel Top & Bottom Line Growth In 21st Century

The Background

Last year, we pitched our deep learning and AI-based fraud protection product, DeepShield.ai, to financial institutions. In doing so, it became clear pretty quickly that our strategy of pitching our product was dead wrong.

Let’s take a step back and look at what led us to pitch our product. Two years ago, we built an MVP payment platform that was targeted for the used item sales market. Soon after soft launching our product, our users started experiencing payment fraud problems. As it turns out, payment fraud is a huge problem in the used item sales market. Further research into payment fraud revealed a systematic problem in all markets.

The Big Problem

As it turns out, almost all major payment methods experience multi-billion dollar fraud losses every year. Credit card fraud resulted in $3.8 Billion losses in 2015, and ACH (transfers between bank accounts) experienced $1.8 Billion losses in 2015. Clearly, the current methodologies to prevent fraud are not working.

To make the problem worse, mutual consensus is that Federal Reserve mandate to make payments faster will inevitably result in 2x-3x more fraud losses.

We talked to major financial institutions, and they all showed genuine interest in solving this problem. One major bank executive even said to us:

“Anything will be better than what we have today considering this infrastructure has not been updated in 30+ years”

Music to our ears…

Our Solution

We got to work and built a kick-ass deep learning and AI-based product that, without human intervention, could predict more than 80% of fraud. Even better, our product required zero operations cost as our product’s self-learning algorithms could automatically protect against new kinds of fraud. We were happy and confident that our product was going to be a huge hit.

The Sales Process

We started pitching our product, DeepShield.ai, to payment guys at various financial institutions.

Demonstrating the problem was easy. Almost everybody already knew the problem existed.

But demonstrating how our solution can help eliminate the problem turned out to be a non-trivial problem. After our pitches and demos, we experienced blank faces and puzzled looks. We knew there was a problem with our pitch. We just didn’t know what it was…

The Key Insight

Then late last year, we got the key insight that helped us understand why our pitch was proving ineffective. We were in a meeting where we presented our product to a very large financial institution. After an hour long session, where we highlighted the benefits of deep learning and how our product is more effective than other solutions, one key executive asked,

“So where will we program the rules in this deep learning product?”

Our jaws dropped to the floor.

If you have a machine learning background, you must already know why… I’ll explain for the rest of the readers.

Deep Learning is Nothing Like Programming

Traditional fraud protection is done using rule based engines. Rule based engines allow financial institutions to program static rules into the system to flag suspicious transactions. E.g. A simple rule might be, if the customer is a customer for less than 365 days, trigger 2 factor authentication for any payment of more than $1000.

Deep learning solutions don’t work like that. There are no rules to program into the system. In contrast, deep learning systems program themselves. There is no need to look at patterns manually and deduce that a certain behavior is associated with high risk of fraud. Also, there is also no need to generate and program a rule in the system to mitigate the losses from a certain behavior.

Instead, deep learning based systems detect patterns and program themselves without any human intervention.

By now, you can probably guess what was the problem. We assumed that product guys at financial institutions at least understood the deep learning paradigm and how it works. That assumption turned out to be false.

The Deep Learning Age

After every couple of decades, computing goes through a complete paradigm shift that renders the old computing model obsolete. However, this shift is never smooth or instantaneous.

Before reaching full acceptance, there are multiple states of confusion, awareness and understanding of the new computing paradigm. At first, there is total unawareness of the new paradigm. Then, there is some awareness with little understanding that gives rise to mass confusion and fears (e.g. machines will take over the world). Next is a long period slow acceptance, where there is a better understanding of the paradigm and there is some acceptance. Finally, after reaching a critical threshold of awareness and understanding, there is universal acceptance of the new computing paradigm.

About a hundred years ago, computation was done through mechanical devices like gears and levers. Since then there have been multiple computing paradigm shifts.

The first paradigm shift occurred through the invention of solid state technology. Mechanical device programmers had a hard time understanding how logic can be programmed by simply printing on a circuit board.

The second paradigm shift occurred through the invention of microprocessor. This time printed circuit board programmers had a hard time understanding how can logic be re-programmed on the same printed circuit board through software.

Today, deep learning is bringing another paradigm shift to the computing world. Even software developers are having a hard time wrapping their head around the concept that a piece of software can program and reprogram itself without even writing another line of code.

We’re At Peak Confusion Again

The Current State of Deep Learning Acceptance

With the advent of deep learning, we are again at peak confusion state. There is more confusion around deep learning today than there was confusion around software in 1980’s. This confusion is a big problem for deep learning adoption.

Building deep learning awareness and understanding is the only way to increase acceptance. Adoption of deep learning based solutions will remain a challenge, unless

  1. Deep learning community does a good job of explaining how deep learning actually works
  2. Deep learning solution providers can clearly demonstrate top and bottom line growth through their solutions

Software revolutionized the world in the late 20th century; I believe deep learning will revolutionize the world in the 21st century in a much more dramatic way. So, I will do a series of posts to show how deep learning can help solve real world problems. My focus will be to demonstrate how deep learning actually works and how enterprises can increase their top and bottom line through deep learning solutions.

Due to the nature of the topic some posts might be more technical than others. Hopefully the accompanying explanation will make the non-technical readers become comfortable with deep learning over time.