Transposed convolution

As explained in Math and Architectures of Deep Learning

A convolution-like operation often described via the transpose of the convolution matrix, used here as a motivation for understanding the matrix form of convolution.

Transposed convolution is a convolution-like operation formed by multiplying with the transpose of a convolution matrix. It performs a kind of upsampling: a lower-dimensional output is distributed back across a larger array in the same proportions used during forward convolution. It is not an inverse, because the forward convolution may discard information when a non-invertible matrix maps more independent inputs to fewer outputs. section 10.5

The operation connects mathematical structure with neural-network reconstruction:

• For a 1D convolution, multiplying by the transposed weight matrix distributes output values back to input positions, resembling proportional blame distribution in backpropagation. section 10.5
• In an autoencoder, convolution and pooling layers encode a high-resolution input into a smaller embedding; decoder layers use transposed convolutions to reconstruct an output with the original scale. section 10.5.1 section 10.5.3
• Unlike nearest-neighbor, bilinear, or bicubic interpolation, transposed convolution learns the mapping during training, distributing values to minimize reconstruction error. section 10.5.3

Output size and a 2D implementation

The transposed-convolution output size follows o′ = (n′ − 1)s + k − 2p. With input size n′ = 3, stride s = 1, kernel size k = 3, and valid padding p = 0, the output size is 5. In two dimensions, a 1 × 1 × 2 × 2 input, a 1 × 1 × 2 × 2 kernel, and stride 2 produce a 1 × 1 × 4 × 4 output. PyTorch implements this with ConvTranspose2d; the demonstrated code fixes the kernel weights and applies it without gradient tracking. section 10.5.2 section 10.5.3

Figure 10.16 1D convolution and its transpose

Use the 1D comparison to inspect how convolution and its transpose distribute values. section 10.5

How Math and Architectures of Deep Learning covers it

How deep it goes

Derives the matrix intuition, output-size formula, tensor shapes, and PyTorch use, but defers autoencoder training details.

Examples

Demonstrates one-dimensional matrix transposition, a 3-to-5 output calculation, and 2D PyTorch upsampling from 2 × 2 to 4 × 4.

What it assumes

Assumes forward convolution as context, then develops the transpose operation through matrices, embeddings, and reconstruction.

What you can do after

Apply transposed convolution via PyTorch ConvTranspose2d and calculate its output dimensions from stride, kernel, and padding.

How it connects

Ties transposed convolution to convolution, backpropagation, autoencoders, embeddings, pooling, and upsampling.

liveBookmanning.com

Read Math and Architectures of Deep Learning on liveBook Get the book at manning.com

Also explained in

• 

  liveBookmanning.com
  Learn Generative AI with PyTorchThe 2D computation performed by a transposed convolutional layer to expand an image or feature map.
• 

  liveBookmanning.com
  Inside Deep Learning: Math, Algorithms, ModelsA convolution-like operation used to increase spatial resolution in a tensor, often to undo downsampling from pooling in dense prediction models.
• 

  liveBookmanning.com
  Build a Text-to-Image Generator (from Scratch)A transposed convolutional layer increases spatial resolution by applying a learned convolution-like operation that can expand spatial dimensions.
• 

  liveBookmanning.com
  Grokking AI Algorithms, Second EditionA transposed convolution is a learnable convolutional operation that projects a lower-resolution representation into a higher-resolution output space.

Compare how 5 books explain transposed convolution →

Related concepts

• up blockused by · An up block is a U-Net component that increases spatial resolution through operations such as transposed convolution.
• upsamplingimplements · Upsampling is a process that increases the number, density, or spatial or temporal resolution of samples or values in a dataset or representation by adding, duplicating, or estimating them.
• convolution kerneluses · A convolution kernel is a small array or tensor of weights applied across local input regions to compute weighted responses and detect or transform patterns.
• convolutional layercontrasts with · A convolutional layer is a neural-network layer that applies shared learned filters to local regions of structured input to produce feature representations.
• decodeused by · Decoding is the process of transforming an encoded or internal representation into an interpretable or target-form output.

This explainer was generated by Manning's knowledge graph from the text of Math and Architectures of Deep Learning. The section links point to the parts of the book it draws on.