AI, ML, and Deep Learning — A Complete Beginner's Guide
If terms like AI, machine learning, and deep learning feel intimidating, you are not alone. Once you understand how they relate to each other, a large portion of the AIF-C01 exam becomes much more approachable.
---
How AI, ML, and DL Relate
Think of Russian nesting dolls. AI is the outermost doll. Inside it sits ML. Inside ML sits DL. Inside DL sits Generative AI.
AI and ML are NOT the same thing. AI is the larger field.
---
Artificial Intelligence (AI)
AI covers any technology that lets computers perform tasks that normally require human intelligence — recognizing faces, detecting fraud, or playing chess.
---
Machine Learning (ML)
Instead of writing explicit rules, ML learns rules from data. Feed it thousands of labeled examples and it finds patterns on its own.
Two key types of output: Regression: predicts a number (house price, temperature) Classification: predicts a category (spam/not spam, dog/cat)
Data split for training:
| Dataset | Ratio | Purpose | |---|---|---| | Training | 60–80% | Model learns | | Validation | 10–20% | Tune hyperparameters | | Test | 10–20% | Final evaluation |
---
Deep Learning (DL)
DL uses multi-layer neural networks inspired by the human brain. More layers = ability to learn more complex patterns. Requires large data and GPUs.
Neural network layers: Input Layer: receives raw data Hidden Layers: extracts features and patterns Output Layer: produces predictions
---
Transformers and LLMs
Transformers process the entire sequence at once (parallel), unlike older RNN models that read word-by-word. The Attention mechanism determines how much each word relates to every other word. BERT (bidirectional) and GPT (generative) are the landmark transformer models.
---
Learning Types Summary
| Type | Data | Example Use | |---|---|---| | Supervised | Labeled | Spam detection | | Unsupervised | Unlabeled | Customer segmentation | | Semi-Supervised | Mixed | When labeling is expensive | | Self-Supervised | Unlabeled (auto-labeled) | GPT, BERT pre-training | | Reinforcement | Reward signals | Game AI, robotics |
---
Key Model Reference
| Model | Primary Use | |---|---| | GPT | Text/code generation | | BERT | Bidirectional text understanding | | GAN | Synthetic data/image generation | | XGBoost | Tabular data classification/regression | | RNN | Sequential data, speech | | ResNet | Image classification |
---
Exam Key Points
AI contains ML contains DL — they are nested, not equal. Transformers process sequences in parallel, not sequentially. Labeled data = supervised learning; unlabeled = unsupervised. GPT and BERT use self-supervised pre-training. Diffusion models add noise (forward) then remove noise (reverse) to generate images. Training/Validation/Test split: 60–80% / 10–20% / 10–20%.