Foreword by Prof. Shireesh B. Kedare, Director, IIT Bombay
Al is already changing how organizations make decisions, serve citizens and customers, manage risk, build products, and compete. In spite of it being common knowledge that Al will reward the prepared, many teams remain stuck in pilots, unsure how to separate real value from noise.
Written for leaders, professionals, and policymakers who need to understand this technology beyond buzzwords and hype, Grounded gives readers a practical foundation in the language, logic, and trade-offs of modern Al. It explains machine learning, generative AI, neural networks, grounding, retrieval augmented generation, Al agents, responsible Al, policy, risk, and implementation in clear, accessible language. The goal is to help decision-makers ask sharper questions: What problem are we solving? Is the data ready? What risks are we accepting? How will we measure success? What should remain human-led?
Grounded also explores Al’s impact across government, education, healthcare, agriculture, manufacturing, telecommunications, climate, and employment, with a special focus on India and developing nations. It examines both sides of the Al opportunity: productivity, access, innovation, and public-service transformation on one side; bias, privacy, misinformation, security, accountability, and job disruption on the other. The book closes with a practical implementation framework for organizations moving from curiosity to action: discovery, data strategy, model selection, prototyping, piloting, change management, governance, operations, audit, and continuous improvement.
Written from the combined perspective of public-sector leadership and enterprise technology implementation, Grounded helps readers cut through the noise, spot the risks, identify the opportunities, and make better decisions in an AI-driven world.
Here is an excerpt from the book:
A Deeper Dive into How Machines Learn
This chapter is the technical heart of the primer, a compact, practical deep dive into how machines learn and how modern AI systems are built and evaluated. It covers the mental models (supervised learning, gradient-based optimization, bias/variance, neural nets), testing and validation practices, and a full “anatomy” of a generative AI system.
If you are an engineer looking to venture into AI, this is where you’ll find the conceptual tools to choose model families, tune training, and design evaluation rigs.
If you are a CXO/business leader: read it to gain the right questions and tradeoffs (data readiness, cost/latency envelopes, explainability needs) so you can judge proposals and risks without getting lost in math.
There is a saying that algebra is the math of relationships, calculus is the math of change and statistics is the math of data.
We need algebra to understand the relationships within the data; we need calculus to help us understand how those relationships within data change with time and context and finally we need statistics to make sense of the estimations and predictions within the range of probabilities.
We need all three to make machine learning work. Along with that we will require basics of how to work with matrices, an understanding of optimization problems and basic computer programming.
We will briefly touch on the ideas of supervised, unsupervised and reinforcement learning earlier, before we dive into supervised learning and Generative AI.
2.1 SUPERVISED LEARNING
Assume that you were to predict the cost of a house in a particular neighborhood, let’s assume that the main driver is the square footage, now if you had a sample of 20 houses with different ranges of that parameters you could plot them in a graph and fit a line through it y = ax + b. This is classic linear regression.
If that was the only relevant parameter to predict the price of a home and you had to guess what a 2500 square footage home would cost, it would be easy to predict that spot on the line.
If we add another parameter called proximity to the city, the chart could look like this image below with 3 dimensions. We have 2 parameters and the equation to fit, the curve becomes y = ax2 + bx + c.
The cost is on Y-axis, the square footage on X and proximity is on Z-axis.
Predicting a point on that plane is still doable, this is called multiple linear regression.
This is easy to compute and solve, now imagine a scenario where there are a hundred or even thousands of parameters? We need to create a model to represent the problem before we can make a prediction based on the model.
Solving that equation itself could take immense computing power and time. Finding the solution closest to the point we are trying to guess can be extremely complex.
While fitting this curve on the chart, we are finding an “Optimum” solution to the equation with all parameters. Curve fitting simply means we are looking to draw a smooth line that goes through or near the dots so that you can see the pattern to guess where a new one goes.
Therefore, fitting a line becomes a game of optimization, you look at the line on your chart and compute the average error and try to minimize the error to maximize accuracy.
2.2 GRADIENT DESCENT: GETTING TO THE RIGHT FIT
Imagine you are on stage as a guitarist about to play a show, you realize that one of your strings is out of tune. To tune that guitar string, you might ask your pianist to play the correct note to use as reference and you start tightening or loosening the guitar string to match the note on the piano.
The goal is to turn the tuning knob to make the string tighter or looser bit by bit so that eventually both the guitar string and the piano note are ringing in unison and you get that matching resonant note.
You start with a single first step in any direction (tighter or loose) and test the match, if we find that it’s even more out of tune then we take a step in the other direction till we finally get close enough.
Learning optimization is at the core of machine learning, and one of the most useful methods called gradient descent works like tuning a guitar.
Essentially this is what we are doing with the training and test data, we are finding a solution that minimizes the error in prediction.
In the real world the “loss function” is our ear telling us how good or bad the move was with each change when tuning the guitar, the optimizer is our hand making a change in the right direction. We do this in a loop, make a guess, test it, and make a change in a loop till we get to “good enough.”
2.2.1 Considerations in Gradient Descent
2.2.1.1 Bigger Steps Versus Smaller Steps
Another similarity is the fact that when the string is out of tune, if you start with only smaller changes, it will take forever and you end up with people leaving the show, if you make too large of changes all the time you may never get to the right spot.
You make bigger changes at first and as you get closer, you start making smaller changes to not “overshoot” the right note. Gradient descent does that too. It’s the tradeoff between cost/speed of optimization and chances of finding a good solution. This is also called the learning rate; this is a model learning parameter.
Local minima: Another way to think about this is, to imaging we are trying to move a ball to the lowest point on a golf course, in this scenario, in the below example the entire plane represents the loss function, the peaks represent high error, and the valleys represent low error.
If the ball started at point A, and we kept rolling toward the lowest point, it would end up in a minimum error at point B. However, we can clearly see that it is not the most optimal solution, in this case it is point D.
However, if we started the ball rolling process at “C”, below we would have found D, this is a problem in optimization where you get stuck in local minima.
There are many approaches to overcome this predicament while some focus on changing the learning rate and adding noise to the data to introduce randomness, others focus on giving the “ball” momentum so that it can roll over local drops, while others focus on trying occasional “full resets” and evaluating multiple parallel runs with different starting points.
While gradient descent or variants of it are likely the approach that 90% of machine learning training systems employ today, it is by no means the only one.
Think of it as a box full of 100 nails and a few dozen screws of different types, the gradient descent hammer takes care of the nails and then there are other tools for the different screw types.
2.2.1.2 Data Normalization
Another key consideration while building up data to train your model is the notion of normalization. Consider a scenario where you have the cost of the house in dollars, square footage, and the number of rooms as parameters for your training data.
When you start the optimization process, it may be off because the dollar parameter would be in millions ($2,450,000) while the number of rooms would be single digits (8), and the square foot would be in the thousands (4,000) this would make the big numbers dominate the loss function very uneven.
This will end up ruining the learning speed as the optimizer would spend a lot of energy chasing the biggest parameter.
The way to overcome that is to “normalize” that data, i.e. subtract the mean and divide by the standard deviation so that they are all relatively of similar size.
This approach levels the playing field so that one parameter doesn’t overpower the others. One such method for normalization is z-score/standardization. Min-max scaling is among other approaches.
Click here for complete details, chapter breakdowns, and author information.
