Machine Learning: How It Works and What It Can Do For Your Business

"What is this Machine Learning thing all about?"

"How do we use it for competitive advantage and greater profitability -- before our competitors do?"

I thought it would be helpful to summarize in a brief article how we respond to those questions from our clients.

What Machine Learning Does: The 50,000 Foot View

Machine Learning Is Very Different Than Traditional Programming

Traditional computer programs are written in a step-by-step, highly prescriptive fashion to find specific answers from a collection of data. Machine Learning, on the other hand, takes in large volumes of data on a particular problem domain and finds pertinent correlations and categories of factors without being told by a programmer how to do so. Machine Learning sifts actionable data from datasets that would be insurmountably large to analyze in such a fashion in human time and extremely difficult and repetitive to program using traditional programming techniques.

How Machine Learning "Learns"

Machine Learning makes these discoveries by being structured analogously to the collection of neurons in the human brain, and, like humans, it "learns" by being trained and corrected when necessary.

Machine Learning (ML) systems can be used in many ways, but the two most common are Supervised Learning and Unsupervised Learning.

• Supervised Learning: Imagine if you fed an ML tool thousands of digitized images labeled as "dogs." That would be Supervised Learning, because you labeled each piece of training data. After that training, if your tool is effective and properly configured, it should then be able to tell you that a picture of a cat is not a picture of a dog. It can't tell you it's a cat unless it's also been trained with thousands of digitized images of cats, but it can tell you it's not a dog.
• Unsupervised Learning: A typical application would be, for example, feeding the ML tool the genetic testing results, family histories, and medical records of a large number of a population of patients to see if there are clusters of that clinical data that may correlate to various diseases and conditions and thus be predictive.

In the rest of this article, I'll describe several concrete, real-world applications of Machine Learning, but I'll share one now, then talk about the "under-the-hood" details of how Machine Learning does what it does, and then talk about a few more applications across various industries.

A Quick First Example

Currently, in addition to my duties as Synaptic Consulting's co-founding partner and CTO, I'm co-leading a joint Graduate / Undergraduate project spanning Lehigh University's Master of Financial Engineering program in the College of Business and Economics and the Computer Science and Business (CSB) program in the College of Engineering and Applied Sciences. We are working toward characterizing "Falling Knives" (stocks that decline and do not recover) using Machine Learning.

The idea is to use a huge set of stock-specific data (daily pricing history and SEC 10-K and 10-Q reports back to the date of the company's IPO as well as various data derived from that such as the velocity of change of price and volume, Net and Gross profit trends, etc, etc) as well as a wealth of overlaying macroeconomic data (GDP, CPI, Fed Funds Rate, etc, etc) as input for a Machine Learning-generated profile of historic Falling Knives. Basically, we'll be saying to our Machine Learning engine, "Here are all the Falling Knives over the last 50 years and a huge volume of related data for each as well as extensive economy-wide data. Tell us which of those factors are pertinent to historic Falling Knives and in which combinations and ratios."

Once we have such a profile (or profiles) we can determine the degree to which a stock on a dip conforms or doesn't conform to the Falling Knife profile(s). Does it conform by, say, only 10% (a very bullish indicator) or by 90% (a very bad indicator indeed)? That promises to be a powerful research tool for both individual and institutional investors. (Some of my other work in Machine Learning and related aspects of Financial Engineering and FinTech is here, here, and here)

More Detail: Machine Learning and Neural Networks

Brains and Computers

Our brains are made up of *a lot* of neurons. Estimates range from 50 Billion to 500 Billion. Each neuron is connected to thousands of others, forming a parallel system that can process many pieces of information simultaneously.

Somewhat analogously, modern processor chips contain tens of millions of transistors, small numbers of which are interconnected to form individual "logic gates" which are in turn interconnected to form the general purpose processor at the heart of our PCs and other computing devices.

Until Machine Learning, that's where the similarity stopped.

Traditional Computer Programs vs. Machine Learning

Before the emergence of Machine Learning, computer programs were essentially "one-off" efforts to solve a single problem or at most a class of problems. When you had a new problem to solve, you wrote a new program.

A Machine Learning tool, on the other hand, is a general purpose problem solver in the way the human brain is. Like the human brain, it "learns" by categorizing and correlating data and drawing parallels.

Machine Learning software is composed of millions of small pieces of code called nodes, the collection of which is called a Neural Network. The Neural Network's nodes are interconnected in a similar fashion to our brains' neurons. The connections between one node and others are numbers called weights which are positive if one node "fires" or "excites" another or negative if it suppresses another node. The more common a piece of data is to the body of data items under analysis -- for example a particular range of the rate of change of the number of shares changing hands when analyzing "Falling Knife" stocks -- the more nodes that fire and at greater weights. This is analogous to how your brain's neurons fire more frequently and vigorously when you recognize a pattern and fit a new piece into a puzzle. On the other hand, using our "Falling Knife" example, the year of incorporation of such stocks varies widely and likely has little to do with their decline, so fewer nodes would fire at any significant weight in response to that factor, thus characterizing that piece of data as likely to be unrelated to the stock's decline. In this fashion, data is categorized and correlated by a Machine Learning tool, forming a sort of "template" describing pertinent similarities among members of the set of data items under analysis.

How a Machine Learning Tool Learns

A Machine Learning tool is first trained, then, on completion of its training, goes on to process data based on that training. It makes a determination as to the degree to which a new set of data under analysis is "like" the data on which it was trained.

Neural networks continually improve using a feedback process called back propagation or "backprop." Backprop is the ongoing practice of comparing a Machine Learning tool's actual output to the output it *should have* produced and using the difference to modify -- "tweak" or "dial-in" -- the weights and connections of the nodes in the tool's underlying Neural Network.

Going back to the example at the beginning of this article, imagine if you fed thousands and thousands of pictures of dogs to a Machine Learning tool to train it, and subsequently it did a great job of telling the difference between dogs and other animals such as cats, bears, lions, tigers, etc, but one day you put in a picture of a pygmy pony and it identified it as a dog. This is where Data Scientists use backprop and similar techniques to tune the Machine Learning tool's Neural Network model where and as necessary.

Machine Learning Across Industries

Earlier in my corporate career and since co-founding Synaptic a decade ago, I've used Machine Learning in a diverse set of solutions spanning Manufacturing, FinTech / Financial Engineering, and Business Intelligence. 

• I first used Machine Learning / Neural Nets in counterfeit currency detection nearly two decades ago. We used various wavelengths of visible and invisible light shone on and through various areas of the surfaces of major world currencies and their counterfeit counterparts to reveal differentiating aspects of the interaction of ink and paper. Tens of millions of dollars worth of world currencies were sampled. We then used Machine Learning to categorize and correlate that vast collection of data. The output was a very large number of patterns precisely differentiating genuine currency from counterfeits that no team of counterfeiting experts could possibly have identified even in the span of a career -- giving us an extraordinary low rate of false positives and false negatives. 
• Some years later at Synaptic, I led an effort using Machine Learning to help U.S. and E.U. banking clients meet Know Your Customer (KYC) regulations in detecting money laundering. Though detecting financial transaction patterns used by criminals such as terrorists, drug cartels, and human traffickers to obscure the sources and destinations of their ill-gotten funds is a very different problem domain than counterfeit currency detection, the process of using Machine Learning tools and techniques was very nearly the same. 
• In our FinTech practice, Synaptic is currently working with a major global credit card issuer on two fronts. First, we are guiding their Blockchain development to enable them to accept cryptocurrencies as payment. Concurrently, we are employing Machine Learning to detect Blockchain attacks such as Consensus and Ledger-based attacks as well as Wallet attacks to ensure payment security. The solution encompasses analyzing a vast volume of data traffic using Machine Learning to find data traffic patterns typical of such attacks.
• In a very different arena, we have in our Manufacturing practice used Machine Learning for several years to characterize failed vs. long-lived parts of the same type to ascertain variations in materials, processes, and equipment that contribute to their Quality or the lack of it. We also help clients use the predictive aspect of Machine Learning in New Product Development to assess the reliability of components based on material and process patterns common to previously developed highly reliable components.

The remarkable thing about Machine Learning, unprecedented in the many branches of Computer Science, is that once you become adept in its use to solve one problem, you know how to solve a broad spectrum of entirely unrelated problems. Moreover, while Data Scientists are necessary for the "tuning" of a Machine Learning tool's Neural Networks, the bulk of work can be done by non-technical stakeholders spanning the gamut of Sales & Marketing, Manufacturing, Finance, and all other business functions.

Questions? I'd be delighted to discuss...

*************************************************************************

This publication contains general information only and is based on the experience and research of Synaptic Consulting practitioners. It is not a substitute for professional advice or services, nor should it be used as a basis for any action that may affect your business. Synaptic Consulting and related entities shall not be responsible for any loss by any person who relies on this publication.

About Synaptic Consulting -- Synaptic Consulting offers Consulting, Training, and Business Process Outsourcing services throughout The Americas and the E.U. through its Project Management, Manufacturing, Product Cost Reduction, and FinTech / Financial Engineering practices. For more service portfolio information, please see www.synapticconsulting.info

About Jeffrey Anthony -- Synaptic co-founding partner, CTO, and Technology Practice Leader Jeffrey Anthony held a variety of technology and technology leadership positions prior to co-founding the firm including Global VP of Engineering for the M+M Mars technology subsidiary MEI, heading Technology in the Americas for German Smart Card manufacturer ORGA Kartensysteme GmbH, and in a variety of both Engineering and Marketing roles at Corby Industries. See more at his Bio Page.

Synaptic. The Results You Want. Sooner.