CAS CASDEEP LEARNING SERIES cas.org.pk ↗

CAS STUDIES · INTERACTIVE LEARNING

Deep learning,
made visible.

Build neural networks from first principles. Read the mathematics, manipulate the parameters, and follow every number through the architecture.

Start the first lesson ↓
YOUR FIRST ARCHITECTURE4–3–2–1
INPUTHIDDEN 1HIDDEN 2OUTPUT

Weighted sums. Sigmoid activations.
One forward pass at a time.

THE SERIES

Interactive lectures

05 lectures available
01
INTERACTIVE LESSONFOUNDATIONS · PART 1

Introduction of MLP Forward Propagation

Understand the multilayer perceptron as a chain of weighted sums and sigmoid activations. Inspect every neuron, edit the parameters, and compute a complete 4–3–2–1 forward pass.

Perceptron intuitionWeights & biasesLayer-wise notationSigmoid lab
02
NEW INTERACTIVE LESSONFOUNDATIONS · PART 2

MLP Backpropagation: Individual Derivatives, Chain Rule & Weight Updates

Walk the 4–3–2–1 network in reverse. Build visual intuition with handwritten derivatives, master the 4 individual derivatives, trace how the chain rule multiplies sensitivities across layers, and update every weight with gradient descent.

✍️ Handwritten derivatives4 individual derivativesChain rule path explorerWeight updates W ← W − η∇W
03
NEW INTERACTIVE LESSONFOUNDATIONS · PART 3

Gradient Descent on a 3-Layer Network: Every Weight, Every Update Rule

Derive all 15 parameter updates of a 2–2–2–1 linear network by hand. Write the complete chain rule for each weight, simplify it step by step, check every gradient with live numbers, and watch the network train.

Why the minus sign worksComplete chain rule notationAll 15 update rulesPick-a-weight explorer
04
NEW INTERACTIVE LESSONFOUNDATIONS · PART 4

Gradient Math & Chain-Rule Intuition: Resolving Every 4–3–2–1 Derivative

See how partial derivatives physically measure frozen-knob slopes before expanding all 26 weight and bias gradients on a 4–3–2–1 linear network step by step—including why Layer 1 splits and sums two paths.

Differentiation intuitionOutput & L2 chainsLayer 1 two-path sumAll 26 finite-diff checked
05
NEW INTERACTIVE LABFOUNDATIONS · PART 5

Gradient Lab: Build an MLP & Trace Every Gradient

Design your own multilayer perceptron neuron by neuron, then click any weight, bias, or neuron to trace its forward value, its gradient, and every chain-rule path to the loss. Step through propagation and watch gradient descent train the network live.

Custom architecturesForward & backward steppingChain-rule path tracerLive training
01See the architecture

Diagrams and calculations stay connected.

02Touch the mathematics

Change real parameters and observe the result.

03Build the intuition

Notation follows understanding—not the other way around.