By: Mark Krupnik, PhD, Founder and CEO, Retalon
Since this article will probably come out during Income tax season, let me start with the following example: Suppose we would like to build a program that calculates income tax for people. According to US federal income tax rules: “For single filers, all income less than $9,875 is subject to a 10% tax rate. Therefore, if you have $9,900 in taxable income, the first $9,875 is subject to the 10% rate and the remaining $25 is subject to the tax rate of the next bracket (12%)”.
This is an example of rules or an algorithm (set of instructions) for a computer.
Let’s look at this from a formal, pragmatic point of view. A computer equipped with this program can achieve the goal (calculate tax) without human help. So technically, this can be classified as Artificial Intelligence.
But is it “cool” enough? No. It’s not. That is why many people would not consider it part of AI. They may say that if we already know how to do a certain thing, then the process cannot be considered real intelligence. This is a phenomena that has become known as “AI Effect”. One of the first references is known as Tesler’s theorem that says: “AI is whatever hasn’t been done yet.”
In the eyes of some people, the cool part of AI is associated with machine learning, and more specifically with deep learning which requires no instructions and utilizes Neural Nets to learn everything by itself, like a human brain.
The reality is that human development is a combination of multiple processes, including both: instructions, and Neural Net training, as well as many other things.
Let’s take another simple example: If you work in a workshop on a complex project, you may need several tools, for instance a hammer, a screwdriver, plyers, etc. Of course, you can make up a task that can be solved by only using a hammer or only screwdriver, but for most real-life projects you will likely need to use various tools in combination to a certain extent.
In the same manner, AI also consists of several “tools” (such as algorithms, supervised and unsupervised machine learning, etc.). Solving a real-life problem requires a combination of these “tools”, and depending on the task, they can be used in different proportions or not used at all.
There are and there will always be situations where each of these methods will be preferred over others.
For example, the tax calculation task described in the beginning of this article will probably not be delegated to machine learning. There are good reasons to it, for example:
– the solution of this problem does not depend on data– the process should be controllable, observable, and 100% accurate (You can’t just be 80% accurate on your income taxes)
However, the task to assess income tax submissions to identify potential fraud is a perfect application for ML technologies.
Equipped with a number of well labelled data inputs (age, gender, address, education, National Occupational Classification code, job title, salary, deductions, calculated tax, last year tax, and many others) and using the same type of information available from millions of other people, ML models can quickly identify “outliers”.
What happens next? The outliers in data are not necessarily all fraud. Data scientists will analyse anomalies and try to understand the reason for these individuals being flagged. It is quite possible that they will find some additional factors that had to be considered (feature engineering), for example a split between tax on salary, and tax on capital gain of investment. In this case, they would probably add an instruction to the computer to split this data set based on income type. At this very moment, we are not dealing with a pure ML model anymore (as the scientists just added an instruction), but rather with a combination of multiple AI tools.
ML is a great technology that can already solve many specific tasks. It will certainly expand to many areas, due to its ability to adapt to change without major effort on a human side.
At the same time, those segments that can be solved using specific instructions and require predictable outcome (financial calculations) or those involving high risk (human life, health, very expensive and risky projects) require more control and if the algorithmic approach can provide it, it will still be used.
For practical reasons, to solve any specific complex problem, the right combination of tools and methods of both types are required.
About the Author:
Mark Krupnik, PhD, is the founder and CEO of Retalon, an award-winning provider of retail AI and predictive analytics solutions for planning, inventory optimization, merchandising, pricing and promotions. Mark is a leading expert on building and delivering state-of-the-art solutions for retailers.