Pointwise convolution
A pointwise convolution uses a 1×1 spatial kernel to mix or transform channel values independently at each spatial position, without combining neighboring positions.
Inside Deep Learning: Math, Algorithms, Models
A convolution with a 1-by-1 spatial kernel used to mix or change channel dimensions without expanding the spatial receptive field.
- How it is used
- Uses 1 × 1 convolutions to mix or change channels in CNN blocks, residual bottlenecks, and U-Net blocks.
- How deep it goes
- Explains the k = 1 receptive-field tradeoff and walks through PyTorch blocks and placements, without deriving convolution arithmetic.
- Examples
- Illustrates channel mixing with edge detectors and tests added information-sharing blocks in a CNN.
- What you can do after
- Build it from scratch with nn.Conv2d to mix or resize channels, and decide when its limited spatial reach suits a block.
Read how Inside Deep Learning: Math, Algorithms, Models explains pointwise convolution →
Math and Architectures of Deep Learning
A convolution with a one-by-one kernel that mixes channels at each spatial location and can reduce dimensionality.
- How it is used
- Deploys pointwise convolutions for channel reduction in Inception and shape alignment in ResNet shortcuts.
- How deep it goes
- Explains architectural roles and PyTorch placement, but defers detailed parameter-count derivations and kernel-level internals.
- Examples
- Uses reduced-channel Inception branches and stride-2 ResNet shortcut projections as concrete architectural examples.
- What you can do after
- Apply it via PyTorch to reduce channels in convolutional branches or align residual shortcut dimensions.
Read how Math and Architectures of Deep Learning explains pointwise convolution →
Also covered in
- Fight Fraud with Machine LearningIn early access (MEAP)
What to know first
- channel dimensionA channel dimension is a dimension of image or tensor data whose positions distinguish separate color components or feature maps.
- convolutional layerA convolutional layer is a neural-network layer that applies shared learned filters to local regions of structured input to produce feature representations.
- feature mapA neural-network feature map is a spatially organized array of activation values representing learned features at different locations or channels.
- convolution kernelA convolution kernel is an array or tensor of weights applied to local regions of structured input to compute convolution responses.
Related concepts
- depthwise separable convolutionpart of · A depthwise separable convolution combines per-channel spatial filtering with a pointwise convolution that mixes channels.
- channel dimensionuses · A channel dimension is a dimension of image or tensor data whose positions distinguish separate color components or feature maps.
- DeepLabV3used by · DeepLabV3 is a convolutional-neural-network-based image segmentation architecture that uses atrous convolution and spatial pyramid pooling to capture information at multiple scales.
- Inception moduleused by · An Inception module is a multi-branch neural-network block that processes inputs in parallel with operations having different receptive fields before merging the resulting features.
- residual connectionused by · A residual connection is a neural-network pathway that carries an earlier representation around one or more transformations so it can be combined with a later computation.
- convolution kerneluses · A convolution kernel is an array or tensor of weights applied to local regions of structured input to compute convolution responses.
Questions about pointwise convolution
What is pointwise convolution?
A pointwise convolution uses a 1×1 spatial kernel to mix or transform channel values independently at each spatial position, without combining neighboring positions.
Which Manning books explain pointwise convolution?
Inside Deep Learning: Math, Algorithms, Models and Math and Architectures of Deep Learning.
What should I know before learning pointwise convolution?
Manning books that teach pointwise convolution usually introduce channel dimension, convolutional layer, feature map and convolution kernel first.
What is pointwise convolution part of?
Pointwise convolution is part of depthwise separable convolution. A depthwise separable convolution combines per-channel spatial filtering with a pointwise convolution that mixes channels.
What is pointwise convolution also called?
Pointwise convolution is also called 1 × 1 convolution.
These guides are generated from the text of Manning books by Manning's knowledge graph and link to the sections they draw on. The full books are available on liveBook and at manning.com.

