Module 3 — Deep Learning Theory·Lesson 10 of 22

3.2 CNNs (conceptual)

Interview-ready field notes

The 20-second answer

CNNs learn small filters that slide over an image, recognizing local patterns while reusing the same weights everywhere.

Core ideas to retain

  • A convolutional filter produces a feature map: early filters find edges, later layers combine them into shapes and objects.
  • Weight sharing cuts parameters and gives translation equivariance: moving an object moves its activation map.
  • Stride controls how far the filter moves; padding preserves borders or output size; pooling/downsampling trades detail for scale and efficiency.
  • A receptive field is the region of the input that can influence one activation; it grows with depth.

Interview / OA rule

Output size per spatial dimension is (N+2PK)/S+1\lfloor(N+2P-K)/S\rfloor+1. State the assumptions about dilation if it appears in an OA question.

One good written resource

CS231n — Convolutional Neural Networks — read this after the video when you want a clearer mental model, not more pages of notes.

Most asked

Interview questions to practise aloud

Each answer is the level of detail expected for a strong fundamentals round.

01

Why are CNNs parameter-efficient?

The same small kernel is reused at every spatial location, unlike a dense layer that needs separate weights for every pixel-to-output connection.

02

What do stride and padding do?

Stride downsamples by skipping positions; padding adds border values so edge information is kept and spatial size can be controlled.

03

Convolution vs pooling?

Convolution learns features with weights. Pooling is usually a fixed local aggregation that reduces spatial resolution.