What is a Neural Network?

This cover sets the frame for the entire topic: a neural network is not a digital brain, it is a stack of simple arithmetic with one nonlinear twist repeated across layers. Most learners arrive carrying the brain metaphor, which makes the subject feel mystical and inaccessible. Replacing 'thinking machine' with 'multiplications and one bent line' immediately demystifies it and gives a concrete object to reason about.

The post that follows is deliberately the 'map' angle. We are not yet justifying why neural networks dominate AI or how training works through backpropagation — we are nailing down what the object IS and the vocabulary around it, so the later posts have solid ground to stand on.

The definition slide reframes the network as a composed function rather than a biological system. The key insight to internalize is that the entire model is built from one tiny repeated unit, and the only exotic ingredient is the nonlinear activation. Everything else is the linear algebra most learners already half-know: multiply, add, repeat.

The phrase 'neural is an analogy' matters and is worth pausing on. The loose inspiration from biological neurons is historically real but practically misleading — real neurons are vastly more complex, and nothing about understanding or using a network requires neuroscience. Treating it as pure math is both more accurate and more empowering.

This slide locks down the single neuron because it is the atomic unit of everything that follows. The formula output = f(w·x + b) is the whole story compressed: a dot product between inputs and weights, a bias added, and an activation applied. If a reader truly absorbs this one line, every architecture later in the series is just a different way of wiring these units together.

The practical takeaway is that each part has a distinct job. The weights decide relative importance of inputs, the bias sets where the unit 'fires,' and the activation is what allows the unit to express anything beyond a straight-line relationship. Naming the three roles separately prevents the common blur where beginners treat the neuron as one opaque blob.

The flow diagram makes the abstract formula concrete by walking a value through the neuron in stages: inputs arrive, get multiplied by weights, the bias shifts the sum, the activation bends it, and the result heads to the next layer. Seeing it as a left-to-right pipeline turns the equation into a process readers can trace by hand.

The final node — 'to next layer' — is the detail that previews stacking. A single neuron is rarely useful alone; its real purpose is to feed the next layer. Visualizing that hand-off here primes the layer concept two slides later and the full network diagram that follows.

The weights-and-bias slide isolates the only things training ever changes. This is a clarifying point because beginners often imagine training as some mysterious global process; in reality it is just the gradual adjustment of these numbers. Everything else — the architecture, the activations, the data — is fixed once you start a run.

The distinction between weights and bias is worth stating plainly. Weights are per-connection and scale inputs; the bias is per-neuron and shifts the activation threshold so a neuron can fire even when all inputs are zero. Without a bias term, every decision boundary would be forced through the origin, which severely limits what the network can represent.

This slide introduces layers as the organizing structure: a layer is simply many neurons computing in parallel on the same inputs. The crucial reframe is that depth comes from stacking these parallel groups, with each hidden layer transforming the previous layer's output into a richer representation. The input layer is your raw features and the output layer is the prediction.

The definition of 'deep' is deliberately deflated here. Deep learning sounds like a separate, advanced field, but it just means more than one hidden layer. Saying so removes intimidation and connects this beginner topic directly to the impressive systems readers have heard about — they are the same machinery, only taller.

The network diagram gives readers a canonical mental image: three inputs feeding a four-neuron hidden layer feeding two outputs, every node connected to every node in the next layer. This fully-connected picture is the default 'shape' of a neural network and the thing most people visualize when the term comes up.

Seeing the explicit wiring reinforces two earlier points at once. Every line is a weight, so the diagram makes the parameter count tangible — even this tiny net has dozens of weights. And the left-to-right flow mirrors the forward pass that post 3 will formalize, so the visual quietly seeds the next post's mechanics.

This slide delivers the conceptual climax of the post: without the nonlinear activation, depth is an illusion. Stacking linear layers — matrix multiplies — collapses mathematically into a single linear layer, so a hundred-layer linear network is no more powerful than one. The activation is what breaks that collapse and lets layers compose into genuinely complex functions.

The importance is hard to overstate. The nonlinearity is the difference between a model that can only draw straight boundaries and one that can carve any shape in the data. This is also why the choice of activation (ReLU, sigmoid, tanh) is a topic in its own right — it is teased here and becomes the next day's subject.

This three-line code slide gives readers something concrete to run before any deeper mechanics arrive. Implementing a neuron with a dot product, a bias add, and a ReLU max makes the formula from earlier tangible — they can change the weights and watch the output move. The point is psychological as much as technical: the core unit really is this small.

The choice of np.maximum(0, z) for ReLU is deliberate. It shows that the famous activation is just 'keep positives, zero out negatives,' which is far less intimidating than the name suggests. Seeing the entire neuron in runnable numpy demystifies the building block that every later architecture is made of.

The comparison slide situates neural networks against the regression models many readers already know, preventing the misconception that they are something wholly alien. A linear or logistic regression is, in fact, a neural network with no hidden layers — one layer of weights drawing a straight boundary. The network generalizes it by stacking layers and adding nonlinearity.

The most important contrast is the last row: regression makes you design the features, while a deep network learns its own. That single difference is the seed of the entire 'why it matters' post that follows, so planting it here gives readers a reason to keep going and a bridge into post 2's central argument.

This recap consolidates the five vocabulary terms into a single screenshot-able reference. Spaced repetition of neuron, weights, bias, layer, and deep across the post is intentional — these words appear constantly in the next four posts and in any neural-network material the reader will ever encounter.

The term doing extra duty is 'neuron,' since its definition (weighted sum + bias + activation) silently contains the other four ideas. Listing them together primes the forward pass in post 3, where these exact terms reappear as the steps of a computation rather than static definitions.

The CTA closes the loop and points forward to post 2's argument. Framing the next post around why this simple math powers vision, language, and modern AI gives readers a concrete reason to continue rather than a generic 'follow me.'

It also sets expectations: post 1 was the what, post 2 is the why. Naming that progression helps readers see the series as a structured course rather than disconnected tips, which is the standard the whole 100-day series holds to.