Tutorials · Chapter D (4/4) · ~10 min

Deep learning

Try it → see it → read → next

Stack neural layers so simple signals can become useful learned representations.

Try yourself

Playground

Layer stack

Deep nets stack features: Edges → Shapes → Objects. Freeze early layers and watch later ones adapt.

EdgesLow-level lines & contrast
ShapesCorners, blobs, parts
ObjectsCats, signs, faces

Recap

What you just did

LayerStackViewer stacked Edges → Shapes → Objects and froze early layers. Depth builds representations; freezing shows what transfer means.

Teach

How it works

A tiny forward pass can be written without a framework:

def relu(value):
    return max(0, value)

x1, x2 = 0.8, 0.3
h1 = relu(1.2 * x1 - 0.5 * x2)
h2 = relu(-0.4 * x1 + 1.1 * x2)
output = 0.7 * h1 + 0.9 * h2
print(output)

Training repeatedly runs two directions:

Forward pass computes a prediction layer by layer
Loss measures the prediction error
Backpropagation assigns credit or blame to weights
Optimizer nudges weights and repeats

Mental model: each layer is a lens; training adjusts every lens until the final picture helps the task.

Use it

When you'd use this

Learning patterns from images, audio, and text
Modeling relationships too complex for one simple rule
Reusing a pretrained network and fine-tuning it

Watch out

Deep models need substantial data, compute, and evaluation. More layers do not automatically improve a small tabular problem. They can also fail opaquely, so compare against a simple baseline before celebrating complexity.

Try next

Try this next

Set one hidden activation to zero and recompute the output. Describe how information from that path disappeared.