What you will learn: AI, machine learning and deep learning
What you will learn in this module. You will be able to explain, in plain words, what artificial intelligence is, how machine learning and deep learning fit inside it, why AI suddenly became practical, how training differs from inference, how a GPU differs from a CPU, and which NVIDIA software does which job. This is the first domain of the NVIDIA AI Infrastructure and Operations Associate exam (Essential AI Knowledge, 38% of the marks). Prerequisites: only basic data center knowledge (what a server, a rack and a network are). No programming or mathematics is needed.
Analogy: a rulebook versus a trainee
Imagine you must teach a new employee to spot fake currency notes. One way is to hand over a thick rulebook: "if the watermark is missing, reject; if the thread is the wrong colour, reject". That is classic software: a human writes every rule. The other way is to sit the trainee at a desk with 10,000 real notes and 10,000 fake ones, each labelled, and let the trainee work out the pattern. After enough examples the trainee spots fakes that no rule ever described. The second way is machine learning: the rules are learned from data instead of being typed in.
Three nested circles
Every deep-learning system is machine learning, and every machine-learning system is AI. The reverse is not true.
| Term | Meaning | Example |
|---|---|---|
| Artificial intelligence (AI) | Any technique that makes software behave in a way we call intelligent | A rule-based chatbot, route planning, a vision system |
| Machine learning (ML) | Software that learns patterns from data instead of following hand-written rules | Fraud scoring, demand forecasting |
| Deep learning (DL) | ML that uses neural networks with many layers; needs large data and much compute | Image recognition, speech recognition |
| Generative AI | Deep-learning models that create new text, images, code or audio | A chat assistant, an image generator |
| Large language model (LLM) | A deep-learning model trained on huge amounts of text to predict the next word piece (token) | The model behind a chat assistant |
The three ways machines learn
- Supervised learning
- Learn from labelled examples (this note is fake, this one is real). Most business ML is supervised.
- Unsupervised learning
- Find structure in data with no labels, for example grouping customers who behave alike.
- Reinforcement learning
- Learn by trial and reward, for example a robot learning to walk.
Why a neural network needs so much hardware
A neural network is a large collection of simple maths units (neurons) arranged in layers. Each neuron multiplies its inputs by numbers called weights, adds them up and passes the result on. Learning means adjusting millions or billions of weights until the answers become right. Because every weight must be updated many times over a huge data set, the work is enormous and highly repetitive. That is exactly the kind of work a GPU does well, which is the subject of a later chapter.
Worked example (made-up numbers). A bank checks 10,000 card payments in a day, and 50 of them are fraud. Rule-based: "flag every payment above 50,000 rupees" catches 20 of the 50 frauds but also flags 300 honest payments, and criminals simply stay under the limit. Machine learning: a model trained on millions of past payments learns combinations (an odd shop, an odd hour, a new device) and catches 44 of the 50 while flagging only 60 honest payments. The ML system is not magic; it is better because it learned from far more examples than any person could turn into rules.
Common mistake. Using AI, ML and DL as if they were the same word. When an exam question asks for the broadest term it wants AI; for the technique that learns from data it wants ML; for many-layer neural networks it wants DL.
Exam trap. "Generative AI" is not a separate technology from deep learning; it is a use of deep learning. Expect a question that asks which term is a subset of which.
The pilot that needed a bigger room
A retail company ran a small ML model that predicted stock for 200 shops on one office server. Management then asked for an image-based system that reads shelf photos from 2,000 shops. The same team found the server far too slow: deep learning on images needed GPUs, fast storage for photos, and a network that could move them. The lesson from the project review was that "AI" is not one workload; moving from ML to deep learning changes the infrastructure you must buy.
Lesson: Know which kind of AI a request really is before you size hardware.
"What is the difference between AI, machine learning and deep learning?"
AI is the broad goal of making software act intelligently. Machine learning is the part of AI where the system learns patterns from data instead of following hand-written rules. Deep learning is the part of ML that uses many-layer neural networks and needs large data and GPU compute; generative AI and large language models are built with it. I would add one example for each, such as a rule-based chatbot, fraud scoring and image recognition.
Key takeaways
- AI contains ML, and ML contains deep learning; generative AI and LLMs are deep learning.
- ML learns rules from data; classic software follows rules a person wrote.
- Supervised, unsupervised and reinforcement learning are the three learning styles.
- A neural network learns by adjusting weights; that is heavy, repetitive maths.
- Deep learning is the reason AI now needs GPU infrastructure.