AIF-C01
Deep Dive
Deep learning is the technique behind modern foundation models. Understanding its basic architecture — without diving into the math — gives you the context you need to reason about generative AI for the rest of the exam.
Contents
An artificial neural network is a computational model organized into layers of interconnected "neurons", loosely inspired by the biological brain. You don't need to understand the underlying math for the exam — just its conceptual structure.
Input layer
Receives the raw data: the pixels of an image, the numeric values of a data row, the tokens of a text. Each input neuron represents one feature of the input.
Hidden layers
Each neuron combines the outputs of the previous layer using weights (how much each input matters) and an activation function that introduces non-linearity. "Deep" learning refers precisely to having many hidden layers stacked together.
Output layer
Produces the final result: a class probability, a numeric value, or — in a language model — a probability distribution over the next possible token.
Training, in one sentence
Training a neural network means iteratively adjusting the weights between neurons so that predictions get closer and closer to the correct answer, by comparing the error and propagating it back through the network (backpropagation) to correct the weights layer by layer.
Launch a training job in SageMaker and watch the loss function decrease epoch after epoch.
+1 more...
| Aspect | Classic ML | Deep Learning |
|---|---|---|
| Feature extraction | Manual — an expert defines which features to use (feature engineering) | Automatic — the network learns its own internal representations from hidden layers |
| Data volume needed | Works reasonably well with small or medium datasets | Requires large datasets to generalize well |
| Compute power | Can be trained on CPU in minutes or hours | Generally requires GPUs and can take hours, days, or weeks |
| Example algorithms | Linear/logistic regression, decision trees, random forest, SVM | Deep neural networks: CNN, RNN, Transformers |
| Interpretability | Easier to explain why a decision was made | Harder to interpret ("black box") |
Deep Learning is a subset of Machine Learning, which in turn is a subset of Artificial Intelligence. All modern generative AI foundation models (GPT, Claude, Llama, Titan) are built with deep learning architectures, specifically Transformers.
The exam requires distinguishing these five terms precisely. The easiest way to remember them is as concentric circles: each one is a more specific subset of the previous one.
Artificial Intelligence (AI)
The broadest field: any technique that lets a machine simulate intelligent behavior — from hand-written "if-then" rules to today's most advanced models.
Machine Learning (ML)
Subset of AI where the system learns patterns from data, instead of following explicitly programmed rules.
Deep Learning
Subset of ML that uses neural networks with many layers to automatically learn representations, without manual feature engineering.
Generative AI (GenAI)
Subset of deep learning focused on generating new content (text, images, audio, code) using foundation models trained on massive volumes of data.
Agentic AI
The most specific level: systems that combine a GenAI foundation model with memory, planning, and access to external tools to act autonomously in pursuit of a goal — making decisions and executing actions without human supervision at every step.
Agentic AI in practice
A chatbot that answers a question is GenAI. A system that you give the goal "book the cheapest flight for these dates", which searches options, compares prices, checks company policies, and completes the booking without a human approving each intermediate step, is Agentic AI. On AWS, Amazon Bedrock Agents is the service that enables this kind of autonomous behavior.
Think of it as concentric circles: AI ⊃ ML ⊃ Deep Learning ⊃ GenAI. Agentic AI isn't an even smaller circle inside GenAI, but a behavior layer built on top of GenAI that adds memory, planning, and tools to act autonomously.
Convolutional Neural Networks (CNN) are the classic architecture for computer vision tasks: image classification, object detection, segmentation.
Core idea
Instead of connecting every pixel to every neuron (computationally infeasible for large images), a CNN applies small filters (kernels) that slide over the image detecting local patterns: edges, textures, shapes. Early layers detect simple patterns (edges, colors); deeper layers combine those patterns into increasingly abstract concepts (eyes, faces, whole objects).
AWS services like Amazon Rekognition use vision models based on this type of architecture, already trained and exposed as an API.
RNN (Recurrent Neural Networks)
Designed for sequential data (text, audio, time series). They process the sequence element by element, keeping an internal "state" or memory that's updated at each step, allowing prior context to influence the processing of the next element.
Key limitation: they process sequentially (slow to train in parallel) and lose context from elements far back in the sequence ("short memory").
Transformers
An architecture introduced in 2017 that replaced RNNs as the standard for language. It uses a mechanism called attention: instead of processing word by word sequentially, it evaluates the relationships between all the words in the sequence simultaneously, "paying more attention" to the most relevant parts of the context for each prediction.
This allows for massively parallel training (leveraging GPUs) and capturing long-range dependencies in text — the foundation of all modern LLMs (GPT, Claude, Titan, Llama).
For the exam
You don't need to know how to implement attention mathematically. You need to recognize that text foundation models (the ones behind Amazon Bedrock) are based on the Transformer architecture, and that this is why they can process long context and generate coherent text.
Deep learning training essentially consists of enormous amounts of matrix operations that can be run in parallel. This determines which type of hardware is most efficient for each task.
CPU (Central Processing Unit)
Few cores, very powerful, optimized for sequential, general-purpose tasks. Suitable for classic ML, data preprocessing, and inference of small models, but inefficient for training large neural networks.
GPU (Graphics Processing Unit)
Thousands of simpler cores designed to run many operations in parallel simultaneously — exactly what the matrix algebra of a neural network requires. It's the standard hardware for training and inferring with foundation models at scale.
On AWS, services like Amazon SageMaker let you select GPU-backed instances (EC2 P and G families) for deep learning training and inference. Amazon Bedrock fully abstracts this decision away: AWS manages the underlying hardware to serve the foundation models, you just consume the API.
Did you get this topic?
Test what you just learned
What key characteristic distinguishes the Transformer architecture from traditional RNNs for processing language?
Put what you learned to the test with original AIF-C01 exam questions.
Start AIF-C01 mock examUnlimited mock exams with the Pro plan
Log in to track your progress.