Level 3: Deep Representations, Images, Sequences, and Transfer

Level 2 showed how a neural network combines numbers, produces intermediate values, and learns by changing parameters. Level 3 asks a different question:

What useful information can those intermediate values keep, change, or throw away?

That question is about representations.

Start with one image

Imagine a tiny picture of a vertical line.

The raw input could be a grid of pixel values. But a later part of a network may not need every pixel separately. It may be more useful to receive a smaller description such as:

vertical-edge strength = high
horizontal-edge strength = low

That new description is a representation: information produced from the original input so later calculations can use it.

A representation is useful only relative to a task. If the task is “vertical or horizontal?”, edge direction may be enough. If the task is “is the line on the left or right?”, a representation that throws away location would be a problem.

This Level follows that same idea across images, ordered sequences, embeddings, and transfer learning.

A few words you will meet

You do not need to memorize these before starting.

Term	Plain meaning in this Level
representation	an internal numerical description of an input used by later computation
convolution	a small pattern detector reused across different locations in an image
pooling	reducing a local group of values to a smaller summary
augmentation	creating a changed training example that should keep the same label meaning
sequence	data where order matters, such as words in a sentence or measurements over time
hidden state	a recurrent model's running numerical summary of earlier sequence information
embedding	a learned vector of numbers used as coordinates for an item such as a token or category
transfer learning	reusing a representation learned on one task as a starting point for another related task

Later you will also meet pretrained model and fine-tuning. A pretrained model has already learned parameters from earlier data. Fine-tuning means allowing some or all of those parameters to change on the new task.

What you will learn

By the end of Level 3, you should be able to:

• explain a representation as an internal description used by later computation;
• compute a small convolution and explain why the same detector can be reused across positions;
• trace an image from pixels to local features to a class decision;
• compare pooling methods and name the detail they discard;
• decide whether an augmentation preserves label meaning;
• explain why sequence order and time alignment matter;
• trace a recurrent hidden state and describe it as compressed memory;
• explain why distant influence can shrink or grow through repeated transformations;
• interpret embeddings as learned coordinates whose geometry depends on training;
• compare training from scratch, frozen transfer, and fine-tuning fairly;
• inspect what a pretrained model expects for input and preprocessing;
• analyze failures by meaningful image or sequence groups rather than one average score.

The learning path

Part 1 — Represent images

L3.1–L3.5 move from the general idea of a representation to convolution, a transparent image classifier, pooling, and augmentation.

Keep asking: What information did this transformation make easier to use, and what information did it remove?

Part 2 — Represent ordered data

L3.6–L3.9 move from sequence order to recurrent memory, long-context limits, and embeddings.

The key shift is that position and history can carry meaning. Two inputs containing the same values can require different predictions when their order differs.

After L3.7 — Recurrent Models and Memory, complete Level 3 Mini Checkpoint — Represent, See, and Remember before continuing.

Part 3 — Reuse and adapt representations

L3.10–L3.14 use notebooks on a real handwritten-digit dataset. You will test transfer learning, inspect what a pretrained model expects, fine-tune it conservatively, compare failure groups, and choose among adaptation strategies.

The word pretrained does not mean “already correct for my task.” It only tells you that some parameters were learned earlier. You still need evidence on the new task.

How to work through the Labs

L3.1–L3.9 use browser Labs. Trace the named intermediate values and make one controlled change at a time.

L3.10–L3.14 use notebooks. There, pay equal attention to:

• which data is used for training and evaluation;
• how the input is prepared;
• which parameters are allowed to change;
• validation evidence;
• how the representation changes;
• which groups of examples fail;
• the settings needed to repeat the run.

A run that finishes without an error is not automatically a good experiment.

What mastery looks like

For each major idea, try to reach three levels:

1. Explain it without relying on a framework name.
2. Trace it through one concrete input, representation, or score.
3. Transfer the reasoning to a changed position, sequence, task, or target-data size.

Read Quick Check feedback even after a correct answer; it should explain why the evidence supports the answer.

Level Project

After L3.14 — Adaptation Review: Choose, Fine-Tune, Explain, complete Transfer Learning Across a Real Dataset.

A strong submission should make the adaptation choice understandable to another learner: use the same target evidence for comparisons, state what was reused and what was changed, and include failures and tradeoffs instead of only headline accuracy.