← Deep Learning Book (Goodfellow, Bengio & Courville)

Learn / Deep Learning Book

09 - Convolutional Networks

Advanced mastery of Chapter 9, Convolutional Networks, covering every named section across book pages 330-372.

Learning path

0%

0 of 5 sections marked complete · about 70 minutes

Learning objectives

What you will be able to explain

  • Compute a valid cross-correlation
  • Three ideas behind convolution
  • Convolution output shape
  • Pooling effects and tradeoffs
  • Convolution as an infinitely strong prior
  • Variants of convolution

Section 01

Compute a valid cross-correlation to Convolution output shape

01

Compute a valid cross-correlation

The valid outputs are [2(1)−2,2(2)−3,2(3)−4]=[0,1,2][2(1)-2, 2(2)-3, 2(3)-4]=[0,1,2], so the middle value is 1. Libraries often call this operation convolution even though it is cross-correlation mathematically.

Guided checkpoint

Using the convention common in neural-network libraries (no kernel reversal), correlate input [1,2,3,4][1,2,3,4] with kernel [2,−1][2,-1]. What is the middle output value?

Source: Chapter 9, section 9.1, pp. 331-335

02

Three ideas behind convolution

These structural assumptions reduce parameters, encode locality, and control how representations respond to transformations.

Guided checkpoint

Match each architectural property to its consequence.

Source: Chapter 9, sections 9.2-9.3, pp. 335-345

03

Convolution output shape

floor((32+2(2)−5)/2)+1=floor(31/2)+1=15+1=16floor((32+2(2)-5)/2)+1=floor(31/2)+1=15+1=16.

Guided checkpoint

For a one-dimensional input of length 32, kernel width 5, zero padding 2 on each side, and stride 2, compute the output length using floor division.

Source: Chapter 9, section 9.5, pp. 347-358

Section 02

Pooling effects and tradeoffs to Variants of convolution

01

Pooling effects and tradeoffs

Pooling exchanges precise location information for robustness and summary. The appropriate neighborhood reflects the transformations the task should ignore.

Guided checkpoint

Judge each statement.

Source: Chapter 9, section 9.3, pp. 339-345

02

Convolution as an infinitely strong prior

Hard-coded connectivity and parameter sharing restrict the hypothesis class more strongly than a soft preference would.

Guided checkpoint

Why does the chapter describe convolution and pooling this way?

Source: Chapter 9, section 9.4, pp. 345-347

03

Variants of convolution

These variants alter resolution, equivariance, border behavior, or the strength of sharing while retaining structured linear maps.

Guided checkpoint

Match each variant to its purpose.

Source: Chapter 9, section 9.5, pp. 347-358

Section 03

Receptive-field growth to Convolution across data types

01

Receptive-field growth

Depth, stride, pooling, and dilation compose local dependencies. A 1x1 spatial kernel still performs a learned channel mixing at each location.

Guided checkpoint

Evaluate each claim.

Source: Chapter 9, section 9.5, pp. 347-358

02

Convolution for structured output

Structured prediction uses convolution to emit spatial maps; downsampling may need offset tricks, upsampling, or multiscale paths to restore dense resolution.

Guided checkpoint

A pixelwise segmentation system needs an output label at every input location. Which design concern follows?

Source: Chapter 9, section 9.6, pp. 358-360

03

Convolution across data types

Ordinary convolution assumes a grid and a meaningful translation operation. Irregular domains need a different notion of neighborhood and sharing.

Guided checkpoint

Match the data to a natural convolutional domain.

Source: Chapter 9, section 9.7, pp. 360-362

Section 04

Ways to compute convolution efficiently to Neuroscientific inspiration

01

Ways to compute convolution efficiently

Convolution's structure supports multiple algorithms; no single implementation is best for every shape and hardware regime.

Guided checkpoint

Select all approaches identified by the chapter.

Source: Chapter 9, section 9.8, pp. 362-363

02

Random or unsupervised convolutional features

The convolutional architecture itself supplies strong locality and invariance biases, though supervised feature learning generally adds task adaptation.

Guided checkpoint

What does their historical success demonstrate?

Source: Chapter 9, section 9.9, pp. 363-364

03

Neuroscientific inspiration

Convolutional networks draw inspiration from visual neuroscience but are engineering models, not detailed simulations of the visual cortex.

Guided checkpoint

Match each term to its role in the chapter's account.

Source: Chapter 9, section 9.10, pp. 364-371

Section 05

Do not confuse equivariance and invariance to CNNs in deep-learning history

01

Do not confuse equivariance and invariance

Equivariance preserves transformation structure; invariance discards it at the output level relevant to the task.

Guided checkpoint

Mark each statement as true or false.

Source: Chapter 9, sections 9.2-9.3, pp. 335-345

02

CNNs in deep-learning history

The historical arc joins longstanding architectural ideas with larger datasets, GPUs, and improved training practice.

Guided checkpoint

Evaluate each historical claim.

Source: Chapter 9, section 9.11, pp. 371-372

Knowledge check

Turn understanding into recall.

The quiz now follows the same concepts in scored form. You can return to this lesson from the quiz whenever a gap appears.

Start the quiz →Back to course