Upsampling

As explained in Grokking AI Algorithms, Second Edition

Upsampling is the process of increasing the spatial resolution of a structured representation such as an image or feature map.

Upsampling increases a feature map’s spatial resolution. In a decoder, it enlarges compressed 4×4 maps to 8×8 maps so the network can reconstruct image details. section 12.19

Two common approaches illustrate the trade-off:

• Nearest-neighbor interpolation duplicates each pixel into a 2×2 area, providing a fast, straightforward way to double resolution. section 12.19
• Transposed convolution applies learned filters to each input pixel, painting larger patches whose overlapping values are added together. section 12.19

From 4×4 to 8×8

Imagine giving a small feature map to an artist. Interpolation copies its marks, while a transposed convolution learns brushstrokes that add coherent detail. A U-Net decoder can then produce sixteen 8×8 feature maps. section 12.19

Figure 12.40 Simplistic upsampling with nearest neighbor interpolation

Use the nearest-neighbor visualization to see how each small-map pixel expands into a 2×2 region. section 12.19

How Grokking AI Algorithms, Second Edition covers it

How deep it goes

Explains pixel duplication, learned patch projection, overlap addition, and a NumPy-style implementation, without covering training details.

Examples

Uses 4×4-to-8×8 feature maps, pixel duplication, an artist analogy, diagrams, and transposed-convolution pseudocode.

What it assumes

Assumes feature maps and convolutional layers while making the resolution-change intuition concrete.

What you can do after

Build it from scratch using array operations for transposed convolution and connect the resulting maps to decoder reconstruction.

How it connects

Connects upsampling with U-Net, transposed convolution, nearest-neighbor interpolation, convolution, and skip connections.

liveBookmanning.com

Read Grokking AI Algorithms, Second Edition on liveBook Get the book at manning.com

Also explained in

• 

  liveBookmanning.com
  Natural Language Processing in Action, Second EditionOversampling the minority class increases the representation of underrepresented classes in training data, often by duplicating or resampling examples.
• 

  liveBookmanning.com
  Evolutionary Deep Learning: Genetic algorithms and neural networksAn upsampling layer increases the spatial resolution of an intermediate feature representation, typically without learning a convolutional filter.
• 

  liveBookmanning.com
  Deep Learning with R, Third EditionUpsampling is the expansion of a representation's spatial resolution by generating or restoring values at additional spatial locations.

Compare how 4 books explain upsampling →

Related concepts

• SMOTEincludes · SMOTE is an oversampling technique that generates synthetic examples for a minority class by interpolating between existing minority-class examples and their nearest neighbors.
• Conv2DTransposeimplemented by · Conv2DTranspose is a neural-network layer that applies a learned two-dimensional transposed convolution to increase the spatial resolution of feature maps.
• class imbalanceaddressed by · Class imbalance is a condition in a classification dataset where classes have substantially unequal numbers of examples.
• transposed convolutionimplemented by · A transposed convolution is a learnable convolution-like operation corresponding to the transpose of a convolution operator and often used to increase spatial resolution.

This explainer was generated by Manning's knowledge graph from the text of Grokking AI Algorithms, Second Edition. The section links point to the parts of the book it draws on.