Back to library

🧠Understand Neural Network Fundamentals

Strip neural networks back to arithmetic — weighted sums, a squash function, and stacking. By the end you'll trace a forward pass with a pencil and design a tabular-problem architecture you can defend choice by choice.

Foundations14 drops~2-week path · 5–8 min/daytechnology

Phase 1A Neuron Is Just Arithmetic

See a neuron as arithmetic, not biology

4 drops
  1. A neuron is a weighted sum and a squash, nothing more

    6 min

    A neuron is a weighted sum and a squash, nothing more

  2. One layer draws lines; two layers draw shapes

    6 min

    One layer draws lines; two layers draw shapes

  3. The squash is what makes the whole thing work

    6 min

    The squash is what makes the whole thing work

  4. A network is a stack of weighted sums separated by squashes

    7 min

    A network is a stack of weighted sums separated by squashes

Phase 2Forward Pass by Hand

Run a forward pass on a 2-3-1 net by hand

5 drops
  1. Lay out the 2-3-1 network as numbers, not boxes

    6 min

    Lay out the 2-3-1 network as numbers, not boxes

  2. Compute z, apply ReLU, write down h

    7 min

    Compute z, apply ReLU, write down h

  3. One last weighted sum gives you the prediction

    6 min

    One last weighted sum gives you the prediction

  4. The loss is one number that says how wrong you are

    6 min

    The loss is one number that says how wrong you are

  5. Backpropagation is the chain rule with a spreadsheet

    7 min

    Backpropagation is the chain rule with a spreadsheet

Phase 3Activations, Depth, and Width

Pick activations, depth, and width with reasons

4 drops
  1. You inherited a model with sigmoid hidden layers and slow training

    7 min

    You inherited a model with sigmoid hidden layers and slow training

  2. Same parameter budget — go wider or deeper?

    7 min

    Same parameter budget — go wider or deeper?

  3. Your friend says any neural net can learn any function

    7 min

    Your friend says any neural net can learn any function

  4. Three teammates, three opinions on hidden layer width

    7 min

    Three teammates, three opinions on hidden layer width

Phase 4Design a Tabular Architecture

Design and defend a tabular-problem architecture

1 drop
  1. Draw the architecture for a real tabular problem and defend it

    12 min

    Draw the architecture for a real tabular problem and defend it

Frequently asked questions

What is a neuron in a neural network, really?
This is covered in the “Understand Neural Network Fundamentals” learning path. Start with daily 5-minute micro-lessons that build from fundamentals to hands-on application.
Why do neural networks need activation functions?
This is covered in the “Understand Neural Network Fundamentals” learning path. Start with daily 5-minute micro-lessons that build from fundamentals to hands-on application.
What's the difference between depth and width in a neural network?
This is covered in the “Understand Neural Network Fundamentals” learning path. Start with daily 5-minute micro-lessons that build from fundamentals to hands-on application.
How does backpropagation actually update the weights?
This is covered in the “Understand Neural Network Fundamentals” learning path. Start with daily 5-minute micro-lessons that build from fundamentals to hands-on application.
How do I choose an architecture for a tabular dataset?
This is covered in the “Understand Neural Network Fundamentals” learning path. Start with daily 5-minute micro-lessons that build from fundamentals to hands-on application.