“Not just GenAI” Leadership Series — Part I

ChatGPT (with its underlying AI model, GPT) and several others like it, are examples of Generative AI (GenAI).  While interest in AI has been increasing over the years, ChatGPT was the harbinger of a sudden spike of interest in AI. Very quickly, thought branches related to AI have penetrated several companies across industries, with novel ideas for AI solutions surfacing across company departments. 

There are several discussions afoot on investments/ROI around AI, and decision makers are expected to have a fair degree of understanding about these areas. 

This is a series of bite sized quick-reads that will equip leadership levels with the requisite information around these and related areas, enabling you to question proposals and make more informed decisions around such investments. 

These posts will take a balanced approach — talking about not just the strengths of these technologies, but also areas to watch out for. These topics would intentionally be presented in a simple discussion-like manner, rather than a dry/abstract third-party tone…Let’s dive in.

“GenAI is the talk of the town. But isn’t that a new kid on the block? What about the ‘other’ AI that was being talked about way before GenAI became famous? Is it obsolete already?”

No, the ‘other’ AI is not obsolete. It’s very much alive, kicking, totally relevant today and would continue to be relevant in the future. The vast majority of AI implementations in the corporate world still falls under the ‘other’ AI. Before understanding GenAI and its counterpart, let’s first look at what exactly ‘AI’ is, with a simple set of examples.

Let’s say you write a software code in your favourite programming language, that goes something like this pseudocode below:

If request = “Multiply” then result = a*b

Else if request = “Add” then result = a+b

When you run the code, the machine is able to easily give you the product/sum of any 2 numbers that you provide, regardless of how big the numbers are, based on the conditions provided.

Is this AI?

Yes. This is an example of ‘rules based AI’.

Is this Machine Learning?

No.

Why?

Because, the machine here is just doing your bidding. Yes, doing it fast, but still just following your instructions. 

What’s Machine Learning (ML)?

When an algorithm learns from the data that you provide, without you explicitly defining the rules, it’s called Machine Learning (ML). And yes, ML is also a part of AI.

So, AI is of 2 types — 

• Rules based, where the rules (aka models) are defined by the human, and the machine simply follows them. 
• ML based, where the model is ‘learnt’ by the machine/algorithm by looking at the patterns in the underlying data.

Generative AI is a type of ML-based AI. So is its counterpart — Discriminative AI (the ‘other’ AI). In order to appreciate the continued relevance of Discriminative AI in your organization, it’s worth taking a quick peek at what these 2 terms mean.

Loan Approval Prediction: Traditional ML

A traditional ML problem statement goes something like this — You are in charge of the loan department of a bank. Your team is responsible for predicting whether any new loan applicant is likely to default on the loan repayment. Over the years, your team has been collecting data on various variables that are likely predictors of whether a person will default or not. Some examples are — employment status, monthly income, number of dependents, other existing loans etc. 

You bring a set of heavy-hitter ML algorithms and let them loose on the data. Each algorithm looks at the data, and identifies patterns in the variables that discriminate between defaulters and non-defaulters.

Understanding Discriminative AI in Algorithmic Solutions

Well, there are different approaches to solve the same problem, packaged in different algorithms. It’s always a good idea to let these different algorithms build their own models (patterns) by looking at the same data. Then use ML model evaluation techniques to see which of these models is the best. 

Regardless of which algorithm triumphs, they all have one common characteristic — they are not too concerned about deeply learning the characteristics of a defaulter. All they care about, is to figure out the easiest way to distinguish between defaulters and non-defaulters, by examining the variables. The focus is on how to draw the boundaries between the 2 classes. Hence this cohort of algorithms are call Discriminative AI (or Discriminative ML, if you will…both are same, since ML is a part of AI).

Comparing Discriminative AI and Generative AI Approaches

GenAI algorithms, on the other hand, want to truly understand the nature of defaulters. For e.g., what is the probability distribution of the ‘monthly income’ variable among the 2 classes — defaulters and non-defaulters. GenAI is not happy with just being able to draw a boundary between the 2 classes. It seeks to have a deeper understanding of the nature of values present in each variable, for a particular class.

And it is this inclination that enables GenAI to actually generate new data — text, images etc — that conforms to the overall nature of the original data, yet is not exactly the same as the original data. 

So, doesn’t GenAI win slam-dunk, since it attempts to understand the data better? Why is Discriminative AI even relevant any more?

The answer to the above questions is founded on where the 2 approaches focus on — Discriminative AI focusses laser sharp on the boundaries between the classes; whereas GenAI focusses on the inner characteristics of the classes themselves. Generally speaking, GenAI can provide some results even when the data is less. Discriminative AI struggles in such ‘small data’ scenarios. On the other hand, if you have a sizeable chunk of clean data, Discriminative AI can provide more accurate results than GenAI for traditional AI use cases like Classification, Regression etc.

Wrap-up

This post examined a fundamental question in many of our minds — What is GenAI, and whether it is the ‘latest’ version of AI superseding all the others before it? The answer is — no. Discriminative AI techniques will continue to be relevant, and both GenAI and Discriminative AI will play tag team to solve some of the vexing problems plaguing enterprises.