Introduction to AI and Machine Learning Concepts

Let's start with the most important term: Artificial Intelligence, or AI. AI is the broadest concept. It means any technique that makes a computer behave in a way that looks intelligent. That could be a chess program that beats a grandmaster, a voice assistant that answers your questions, or a system that predicts which customers will cancel their subscription. AI is the umbrella term. Under that umbrella sit many subfields, and machine learning (ML) is the most important one.

Machine learning is a subcategory of AI. Instead of a programmer writing explicit rules for every possible situation (like 'if the cake is pale, bake for five more minutes'), ML uses algorithms that learn patterns from data. An algorithm is simply a step-by-step procedure for solving a problem—like a recipe, but for the computer. In ML, the algorithm analyses a large set of examples, called a training dataset, and finds statistical relationships. For example, an ML algorithm might look at thousands of emails labelled 'spam' and 'not spam'. It learns that emails containing the word 'free' are more likely to be spam. It never sees the rule 'if the word free appears, mark as spam'. It discovers that rule by itself.

Now, deep learning is a specialised type of machine learning. Think of it as ML on steroids. It uses something called neural networks, which are computer systems loosely inspired by how neurons in the human brain work. A neural network has layers of interconnected 'nodes'. In a simple neural network, you might have an input layer (where you feed in data, like the pixels of an image), one or two hidden layers (where calculations happen), and an output layer (where the result comes out, like 'cat' or 'dog'). Deep learning means the network has many hidden layers—sometimes hundreds. These deep networks can learn extremely complex patterns, which is why they power things like facial recognition, speech-to-text, and self-driving cars. But they also require vast amounts of data and computing power.

Next, generative AI. This is a type of deep learning that creates new content. 'Generative' means it generates something that did not exist before. A traditional ML model might look at a photo and say 'this is a cat'. That is discriminative—it is sorting or classifying. A generative AI model, like the one that powers ChatGPT or DALL-E, creates a new cat picture or writes a new poem about a cat. It has learned the statistical patterns of what cats look like from millions of cat photos, and then it can assemble pixels into a completely new image of a cat that has never existed. Generative AI models are typically very large neural networks trained on enormous datasets. They are remarkable at producing human-like text, photorealistic images, music, and even computer code.

Finally, MLOps. This is 'Machine Learning Operations'. It is the set of practises for deploying, monitoring, and maintaining ML models in the real world. Building a model is only half the job. Once it is built, you have to put it into a production system where it can make predictions on live data. That production system needs to track the model's performance over time. Data changes (this is called 'data drift') and the model might become less accurate. MLOps is the discipline of managing that whole lifecycle: versioning the data, versioning the model, automating the retraining process, and ensuring the model remains reliable and secure. It borrows concepts from DevOps (software development and IT operations), but it is tailored for the unique challenges of machine learning.

To summarise the hierarchy: AI is the broadest field. It includes traditional rule-based systems and ML. ML is a subset of AI. Deep learning is a subset of ML. Generative AI is a specific application of deep learning focused on creating new content. MLOps is the operational practises that keep ML systems running smoothly in production.

Imagine you work as an IT professional for a mid-sized e-commerce company that sells customised sneakers. The company has been doing well, but the marketing team wants to use AI to predict which customers are likely to buy a new 'vintage high-top' design. They want to send targeted emails, not mass emails.

Your first task is to choose the right AI approach. You are not going to write a traditional program with if-then rules (like 'if customer age is 18-25 and has bought sneakers before, send email'). That would be too rigid and would miss patterns you cannot guess in advance. Instead, you decide to use machine learning. You pull historical data from the company's database: past purchases, browsing history, time spent on the sneaker page, whether they clicked previous email offers, their location, and their age. This data becomes the training dataset. You use an OCI service called OCI Data Science to build a machine learning model. The model learns which combinations of those features are most predictive of a future purchase.

But the marketing team also wants a personalised email for each customer. They do not just want a yes/no prediction; they want a creative subject line and product description that appeals to that specific person. For that, you turn to generative AI. You use OCI Generative AI Service to access a large language model (LLM). You feed it the customer's browsing history and past purchases, and you ask it: 'Write a short, friendly email subject line and two-sentence product description about our vintage high-top sneaker, aimed at a customer who previously bought leather boots.' The LLM generates unique text for each of the ten thousand customers.

Now, you have a machine learning model that predicts who will buy, and a generative AI model that creates the email text. This is where MLOps becomes critical. You cannot just run these models once. You deploy the prediction model onto OCI using OCI Model Deployment. You set up a monitoring dashboard that tracks the accuracy of the model every week. After one month, you notice that the model's accuracy is dropping. Customers who were predicted as 'likely to buy' are actually buying less often. You investigate and discover a concept called 'data drift'. A new competitor has entered the market, and customer behaviour has shifted. Your model was trained on old behaviour.

With MLOps, you have automated pipelines set up using OCI Data Pipeline and OCI Events. When the monitoring system detects that accuracy has fallen below a threshold, it triggers a retraining job. The retraining job pulls the latest three months of data, retrains the model, and validates it against a holdout dataset. Once the new model passes validation, it is automatically promoted to production, replacing the old one. This whole process—monitoring, detecting drift, retraining, redeploying—is MLOps in action. Without it, your model would have become useless within weeks. The IT professional in this scenario does not just build a model; they build a system that keeps the model healthy over time.