Deep Learning

Classical machine learning needs you to tell it what to look at — "consider the house's size and location." Deep learning's superpower is that it figures out what to look at by itself. Show it enough photos and it discovers, on its own, that edges make shapes, shapes make eyes, and eyes make faces. This course explains how — without the heavy maths.

The name sounds intimidating, but the idea is gentle. A neural network is built from tiny units called neurons. Each neuron does something trivial: take some numbers in, multiply each by a "weight" (how much it cares about that input), add them up, and pass the result on.

One neuron is almost useless. But connect thousands into layers — the output of one layer feeding the next — and something remarkable happens. The network can represent astonishingly complex patterns.

Remember from the maths course: the inputs are a vector, each layer is a matrix that transforms it, and learning the weights is gradient descent walking downhill. A neural network is exactly those three ideas, stacked and repeated. Nothing more mysterious than that.

Here is the insight that makes deep learning special. Each layer learns to spot bigger ideas built from the layer below it. You never tell it what to look for — it discovers a hierarchy on its own.

For a face-recognition network, researchers have actually looked inside and seen this happen:

  Layer 1  →  edges and simple lines        /  \  |  —
  Layer 2  →  corners, curves, textures     ◜ ◝  ⌒
  Layer 3  →  parts: an eye, a nose, an ear  👁  👃
  Layer 4  →  whole faces                    🙂

This is the difference from classical ML. There, you had to hand-pick the features ("measure the distance between the eyes"). Deep learning learns the features itself, which is why it took over for messy, real-world data like images, audio, and language — where nobody can list all the right features by hand.

A CNN (Convolutional Neural Network) is the design built for images. Its trick is to scan the picture with small "filters" — little windows that slide across the image looking for one thing each (a vertical edge, a patch of red, a curve).

Think of it like looking at a large painting through a small magnifying glass, sweeping it across the canvas, noting where you find each feature. Early filters find edges; later layers combine them into shapes and objects — the hierarchy from the last chapter.

Where CNNs are used every day:

• Your phone's camera finding faces to focus on
• Medical imaging flagging a suspicious area on a scan for a radiologist to check
• A factory line spotting a cracked part
• Self-driving cars recognising pedestrians and signs

Images are static, but lots of data comes in order: words in a sentence, notes in a melody, days of stock prices. For these, what came before changes the meaning of what comes next. "I grew up in France, so I speak fluent ___" — to fill the blank you must remember "France" from earlier.

A RNN (Recurrent Neural Network) handles this by reading one item at a time while keeping a little memory of what it has seen so far. It reads "I," updates its memory, reads "grew," updates again, and so on — so by the blank, it still "remembers" France.

RNNs powered early translation, speech recognition, and text prediction. But they have a weakness: they struggle to remember things from long ago in a sequence — the memory fades. That weakness is exactly the problem the next course (Attention & Transformers) was invented to solve.

Training a deep network from scratch needs enormous data and compute — out of reach for most projects. Transfer learning is the brilliant shortcut: take a network someone already trained on millions of images, and reuse it for your task.

The analogy: hiring someone who already knows how to see, then teaching them only your specific job. The lower layers (which learned edges and shapes) are universal — every vision task needs them. You keep those and only retrain the last bit for your specific labels.