Calculus Basics

In simple terms

Calculus is the study of how things change. The derivative answers “how fast is this changing right now, and in which direction?” — the slope of a curve at a point. The integral answers the reverse: “if I add up all these tiny changes, how much accumulates in total?” — the area under a curve. Most of computing’s use of calculus is about the first one: finding which way is downhill.

More detail

The derivative of a function f(x) is its instantaneous rate of change, written f'(x) or df/dx. Geometrically it’s the slope of the tangent line; where the slope is zero, the function has a peak, valley, or plateau — which is exactly where optima live.

The pieces that matter for computing:

• The chain rule — how to differentiate a function built from nested functions. This is the mathematical core of backpropagation: a neural network is one big composed function, and the chain rule propagates error gradients back through every layer.
• Gradients — for a function of many variables, the gradient is the vector of all its partial derivatives. It points in the direction of steepest increase, so its negative points downhill toward a minimum.
• Optimisation — minimising a loss function by repeatedly stepping in the direction the gradient says reduces it fastest. That single idea, gradient descent, trains almost every modern model.
• Integrals — accumulation: total probability under a distribution, total signal energy, expected values.

You rarely compute these by hand in practice; frameworks perform automatic differentiation. But understanding what a gradient is explains why training works, why learning rates matter, and why models get stuck.

Why it matters

Training a machine-learning model is an optimisation problem, and optimisation is calculus. Every time a model improves during training, a gradient told it which way to nudge millions of parameters. Calculus also underlies physics simulation, signal processing, computer graphics (curves and surfaces), and control systems — anywhere a quantity varies continuously.

Real-world examples

• Neural-network training computes the gradient of a loss with respect to every weight and steps downhill — pure calculus at scale.
• Physics engines integrate acceleration into velocity and velocity into position, frame by frame.
• A camera’s autofocus hunts for the lens position that maximises image sharpness — finding where a derivative is zero.
• Animation eases motion using smooth curves whose slopes are tuned for natural acceleration.

Common misconceptions

• “You need to be fluent in symbolic calculus to do ML.” Frameworks differentiate automatically; what you need is intuition for gradients, slopes, and minima.
• “Gradient descent always finds the best answer.” It finds a local minimum along the path it took; the global optimum may lie elsewhere, which is why initialisation and step size matter.

Learn next

Pair the gradient idea with linear algebra to see how whole layers are differentiated at once, then follow it straight into gradient descent, the algorithm that turns calculus into learning.