Neural Networks 101

Discover the structural building blocks of deep learning—and learn how they fit together in practice.

What you’ll learn

  • Build the perceptron and extend it to multi-layer networks (MLPs).
  • Choose and use activation functions (ReLU, tanh, sigmoid, GELU—when and why).
  • Match loss functions to tasks (MSE, binary cross-entropy, softmax cross-entropy).
  • Implement backpropagation via the chain rule.
  • How nonlinearity gives networks representational power (XOR and beyond).
  • How loss + activation pairings shape gradients (e.g., softmax + cross-entropy).

Hands-on application

  • Implement a 2-layer MLP from scratch (NumPy): forward pass, backward pass, and updates.
  • Rebuild the same model in PyTorch, writing a minimal training loop and comparing to your scratch version.

Prerequisites

  • Comfort with derivatives & the chain rule, vectors/matrices.
  • Familiarity with logistic/linear regression and cross-entropy/MSE.
  • Python experience (NumPy; PyTorch optional but introduced).

Who it’s for

  • Developers who know basic ML and want a rigorous, code-first path into deep learning.
  • Learners who prefer understanding the internals before stacking layers.

Format

  • Duration: ~6–8 hours, fully self-paced.
  • Structure: 6 modules → brief lesson, worked derivation, coding notebook.