Chapter 9.7: Data Types

Deep Learning

CNN

Data Types

Variable Input

Author

Chao Ma

Published

November 29, 2025

Overview

Convolutional networks can operate on many different kinds of data, depending on the number of channels and the dimensionality of the input.

Input Formats

Table 9.1 illustrates how 1D, 2D, and 3D inputs can each appear in both single-channel and multi-channel forms.

CNN input formats across dimensions and channels Figure: Examples of different data types CNNs can process. 1D inputs include audio waveforms and time series. 2D inputs include grayscale images (single-channel) and RGB images (multi-channel). 3D inputs correspond to volumes such as CT scans (single-channel) or videos and skeleton animations (multi-channel spatiotemporal data).

1D inputs include audio waveforms, time series, or multi-sensor signals.
2D inputs include grayscale images (single-channel) and RGB or multispectral images (multi-channel).
3D inputs correspond to volumes such as CT scans (single-channel) or videos and skeleton animations (multi-channel spatiotemporal data).

Flexibility of Convolution

One advantage of convolution is that it naturally supports these different input structures without requiring a fixed dimensionality or a fixed number of channels.

The same convolutional kernel can be applied across various data types because convolution is fundamentally a translation-equivariant, local operation.

Variable-Sized Inputs

Additionally, convolution can also process variable-sized inputs.

As long as the kernel size is fixed, the convolution operation slides across the input regardless of its spatial extent, producing outputs whose dimensions scale accordingly.

This flexibility enables CNNs to handle inputs with different widths or heights—a property that fully connected networks do not have.

Variable-sized input handling Figure: CNNs can process inputs of different sizes. The same convolutional kernel slides across inputs regardless of spatial extent, producing outputs that scale with the input dimensions. This flexibility is unique to convolutional architectures—fully connected networks require fixed-size inputs.

Key Insight

The power of CNNs lies in their flexibility. Unlike fully connected networks that require fixed input dimensions, convolutional networks can handle varying channel counts, different spatial dimensionalities (1D, 2D, 3D), and variable-sized inputs—all with the same kernel. This versatility makes CNNs applicable to a wide range of domains beyond computer vision, from audio processing to volumetric medical imaging.

--- title: "Chapter 9.7: Data Types" author: "Chao Ma" date: "2025-11-29" categories: [Deep Learning, CNN, Data Types, Variable Input] --- ## Overview Convolutional networks can operate on many different kinds of data, depending on the number of **channels** and the **dimensionality** of the input. ## Input Formats Table 9.1 illustrates how 1D, 2D, and 3D inputs can each appear in both single-channel and multi-channel forms. ![CNN input formats across dimensions and channels](cnn_input_formats.png) *Figure: Examples of different data types CNNs can process. 1D inputs include audio waveforms and time series. 2D inputs include grayscale images (single-channel) and RGB images (multi-channel). 3D inputs correspond to volumes such as CT scans (single-channel) or videos and skeleton animations (multi-channel spatiotemporal data).* - **1D inputs** include audio waveforms, time series, or multi-sensor signals. - **2D inputs** include grayscale images (single-channel) and RGB or multispectral images (multi-channel). - **3D inputs** correspond to volumes such as CT scans (single-channel) or videos and skeleton animations (multi-channel spatiotemporal data). ## Flexibility of Convolution One advantage of convolution is that it naturally supports these different input structures without requiring a fixed dimensionality or a fixed number of channels. The same convolutional kernel can be applied across various data types because convolution is fundamentally a translation-equivariant, local operation. ## Variable-Sized Inputs Additionally, convolution can also process **variable-sized inputs**. As long as the kernel size is fixed, the convolution operation slides across the input regardless of its spatial extent, producing outputs whose dimensions scale accordingly. This flexibility enables CNNs to handle inputs with different widths or heights—a property that fully connected networks do not have. ![Variable-sized input handling](cnn_handle_varying_size.png) *Figure: CNNs can process inputs of different sizes. The same convolutional kernel slides across inputs regardless of spatial extent, producing outputs that scale with the input dimensions. This flexibility is unique to convolutional architectures—fully connected networks require fixed-size inputs.* ## Key Insight **The power of CNNs lies in their flexibility.** Unlike fully connected networks that require fixed input dimensions, convolutional networks can handle varying channel counts, different spatial dimensionalities (1D, 2D, 3D), and variable-sized inputs—all with the same kernel. This versatility makes CNNs applicable to a wide range of domains beyond computer vision, from audio processing to volumetric medical imaging.