Cheatsheet: Seeing the field whole

The whole field in three moves

1. depth + data + compute unlocked it        (Lesson 1)
2. ONE engine, wired four ways                (Lessons 2-8)
3. bounded by four honest limits              (Lesson 9)

One engine, four problem shapes

Shape	Wiring	Core idea
Sequences	recurrence → attention	carry or weigh info across positions
Images	convolution	slide small shared filters over patches
Generation	VAE / GAN / diffusion	learn the data’s shape, make new examples
Decisions	reinforcement learning	act, get rewarded, improve a policy

Same neurons + weights from the previous track; only the arrangement changes.

The deep unities (what makes it one field)

• Reuse weights across positions powered both recurrence (per step) and convolution (per patch).
• Learn the shape of the data unified all three generative models.
• One trainer underneath everything: define “wrong,” then gradient descent + backprop nudge the weights until it is less wrong, for vision, language, generation, and decisions alike.

The four limits (always present)

Data hunger · brittleness (OOD/adversarial) · data-slant (bias) · no guarantees + opacity (incl. hallucination). Capability and limitation are the same machinery from two sides.

What deep learning is, now

The neural network from the previous track, scaled with depth/data/compute and arranged to fit the problem’s shape. A single learnable pattern-matching engine, pointed many directions, powerful and bounded. Not magic, not a mind, not nothing.

Place any system in four questions

1. What shape of problem? (sequence / image / generation / decision)
2. How is the engine wired for it?
3. What was it trained on?
4. Where will it break?

Where to go next

• Track 5 (Transformers and LLMs): the depth behind L3’s “attention in brief”, how language models actually work.
• Track 13 (Build Neural Networks from Scratch): hands-on, build it in code.
• Neural Network Intuition (previous track): the engine this survey assumed; firm it up if needed.
• Each problem shape has its own deeper road when one pulls at you.

The durable map

Specific models change in months (new names, bigger numbers). The structure, four shapes, one engine, four limits, lasts. You learned the part that lasts.

The one-line version

Deep learning is one learnable pattern-matching engine, pointed many ways, powerful and bounded; hold that and you can place almost any AI system you meet.