Synapses and silicon.
"Deep learning" sounds like a locked room only mathematicians can enter. It isn't. A neural network is a stack of tiny dials that pass signals to each other and slowly tune themselves to be less wrong. This class turns that into a playground you can poke — no code, no maths anxiety, just a network you can light up with your own hands.
Read a neuron
Explain inputs, weights, bias and activation in plain words.
See learning
Describe how a network gets "less wrong" over time.
Demystify depth
Say what "layers" and "deep" actually mean.
True or false?
"A neural network literally works like a human brain — it has real neurons that think." Decide, then confirm.
The neuron
A neuron is not a brain cell. It's a tiny piece of arithmetic. It takes a few inputs, multiplies each by a weight (how much that input matters), adds a bias (a nudge), and passes the total through an activation function that decides whether to "fire" and how strongly.
output = activation( input₁·weight₁ + input₂·weight₂ + … + bias ) — that's the whole atom of deep learning. Stack millions of them and you get something that can recognise a face or finish your sentence.
The signals
Numbers coming in — pixels, words-as-numbers, sensor readings.
What matters
A learned multiplier on each input. Big weight = strong influence; near-zero = ignored.
The nudge
A constant that shifts how easily the neuron fires.
The decision
A non-linear gate (ReLU, sigmoid). Without it, stacking layers would collapse to one straight line.
Layers & features
Wire neurons into layers and feed one layer's output into the next. The input layer takes raw data, hidden layers transform it into increasingly useful representations, and the output layer gives the answer. "Deep" just means more than a couple of hidden layers.
Edges
In a vision network, the first layers learn simple things — lines, edges, corners.
Parts
They combine edges into textures and parts — an eye, a wheel, a letter.
Whole concepts
They assemble parts into whole objects — a face, a cat, a stop sign.
Nobody hand-writes those features. The network discovers the useful intermediate concepts on its own. That self-discovery is why deep nets beat hand-built rules.
The living network
Toggle the inputs, then fire a signal and watch it ripple through the layers. The output node that lights brightest is the network's "answer". Re-roll the weights to see how the same inputs can lead somewhere completely different — that's why training (finding good weights) is everything.
How it learns
Training is a feedback loop. The network guesses (the forward pass), a loss function measures how wrong it was, backpropagation works out how much each weight contributed to the error, and an optimizer nudges every weight a little to reduce it. Repeat millions of times and the errors shrink.
Loss
Defines what "wrong" means — the distance from the right answer.
Gradient descent
The loss is a landscape; the gradient points downhill. The optimizer steps the weights toward lower error.
Backpropagation
Assigns each weight its share of the blame so it can be corrected.
Learning rate
How big each step is. Too big overshoots; too small crawls.
The aim isn't to memorise the training set; it's to generalise to new data. Memorising noise that doesn't transfer is overfitting. Bigger ≠ automatically better.
The helpline triage
A queer youth helpline gets more messages than volunteers can read in real time. They train a small network to flag urgent messages first.
Words → numbers
Each message becomes numbers. The network learns which patterns (certain words, intensity, time of night) correlate with "needs a human now."
It ranks, people decide
The net only re-orders the queue. A trained volunteer still reads and responds — the model is a triage aid, never the final word.
A network is only as fair as its data. If past messages skewed toward one dialect, the model may under-rank others — so the team audits its misses and keeps a human reading every flagged-low message in a crisis context.
Play with a real one
Visual playgrounds let you train a tiny network in your browser and watch the decision boundary form. Go find one and note what changes when you add layers.
What to look up
"TensorFlow Playground" · "neural network visualizer" · "activation function ReLU vs sigmoid".
Explain it to a friend
In 3–4 sentences, explain a neural network to someone who's never coded — no jargon, just a clear mental model. Use a metaphor if it helps.
Model answer — tap to compare
"A neural network is a stack of simple units that each weigh their inputs and pass on a signal. We don't program the rules — we show it many labelled examples, measure how wrong its guesses are, and automatically nudge the weights to be less wrong. Over time it discovers useful patterns by itself. That's why it's powerful, and also why it fails quietly: it only knows what its data showed it."
Quiz — 5 questions
Common mistakes
"It thinks like a brain"
It's inspired by brains, but it's arithmetic — no understanding, no intent. Fix: picture dials and signals, not consciousness.
"Bigger always wins"
More layers can overfit and cost more without helping. Fix: match size to the problem and watch generalisation, not raw size.
"The output is objective"
A net only knows its training data. Fix: assume its blind spots mirror the data's gaps.
"Accuracy is the whole story"
High average accuracy can hide bad failures on a minority. Fix: check where it fails, not just how often it succeeds.
Dials, not magic
Now that a network is just weighted signals being tuned to be less wrong, does AI feel more controllable — or more fragile? Where would you refuse to let a statistical guess make the final call?
You've opened the black box and found dials, not magic. Next, we point this machinery at pixels and learn how diffusion models turn pure noise into images.