# Convolution: Where Modern AI Begins > Source: > Published: 2026-08-14 06:39:50+00:00 Every day, billions of digital interactions rely on an algorithm you might never have heard of. When Spotify smooths out audio signals, when your phone’s camera portrait mode selectively blurs a background, or when an autonomous vehicle identifies a pedestrian in real time — the exact same core mechanism is working behind the scenes. It’s called **convolution**. Convolution is often introduced through integral equations that obscure its core mechanics. In this guide, we will clarify the operation step by step. Imagine a 1D signal representing daily temperature readings over a week. To extract a trend and reduce daily noise, you apply a moving average. You slide a small window across your data, multiply the values inside that window by fixed weights, and sum them up. That moving window operation is convolution in 1D. Mathematically, the discrete convolution of a continuous or discrete signal f with a filter (or kernel) g over a discrete domain is defined as: In practical machine learning contexts, signals and kernels have finite lengths. If f is an input vector of length N and g is a kernel of length M, the convolution value at index n is calculated as: Note on Signal Flipping:Formally, pure mathematical convolution flips the kernel g (hence g[n-m]). However, in deep learning frameworks like PyTorch and TensorFlow, the kernel is used directly without flipping — an operation technically known ascross-correlation. Because the weights of the kernel are learned during training, the distinction is functionally irrelevant in practice, and both operations are referred to as convolution. Consider a simple 1D signal f and a smoothing kernel g: Applying g to f computes a localized 3-element moving average. At index n=3 (where the signal spikes to 8): The sharp spike at 8 is smoothed out relative to its neighbors. When moving from 1D to 2D, the fundamental mechanics remain unchanged. Instead of sliding a 1D window along a line (left-to-right), a 2D window (kernel) slides across a 2D grid (height and width) in two directions: left-to-right and top-to-bottom. In digital image processing, an image is a 2D matrix of pixel intensities I(x, y). A 2D kernel K is a small matrix — typically of odd dimensions like 3 x 3 or 5 x 5. For an image I and a kernel K of size M x N, the 2D discrete convolution at position (i, j) is: Or, using the un-flipped cross-correlation formulation standard in deep learning: The spatial output size O for an input of size W, kernel size K, padding P, and stride S is calculated via: Before deep learning, engineers hand-crafted kernels to extract visual features. The Vertical Sobel Kernel K_y highlights vertical edges by computing horizontal intensity gradients: When placed over a region with uniform brightness, the negative and positive weights balance out to 0. When placed over a vertical transition (a sudden change from black to white), the output yields a high absolute magnitude — registering a detected edge. To process volumetric or temporal data, we scale the spatial kernel across three axes (X, Y, Z). The 3D convolution formula evaluates a 3D kernel K sliding across a 3D input volume V: RGB Image vs. 3D Convolution:A standard color image has shape H x W x 3, where 3 represents color channels. Standard operations on this format use2D convolutions with multi-channel inputs, not 3D convolutions. A 2D convolution kernel for RGB has dimensions 3 x 3 x 3, but it slides only in2 directions(height and width). In a true 3D convolution, the kernel slides in3 directions. Mathematically, convolution extends to k dimensions. For an input tensor T and kernel K in k-dimensional space, the value at multi-index is expressed as: where m = (m_1, m_2, …, m_k) indexes over the kernel volume Omega. Standard Fully-Connected layers scale poorly with high-dimensional spatial data. For a modest 1000 x 1000 pixel RGB image, a single hidden node in a fully connected layer requires 3,000,000 weight parameters. Convolutional networks solve this through three core principles: ``` python import torchimport torch.nn as nn# 1D Convolution: Audio signal or time series# Input shape: (Batch Size, Channels, Length)x_1d = torch.randn(1, 1, 100)conv_1d = nn.Conv1d(in_channels=1, out_channels=16, kernel_size=3, stride=1, padding=1)output_1d = conv_1d(x_1d)print(f"1D Output Shape: {output_1d.shape}") # Shape: (1, 16, 100)# 2D Convolution: Single-channel image# Input shape: (Batch Size, Channels, Height, Width)x_2d = torch.randn(1, 1, 64, 64)conv_2d = nn.Conv2d(in_channels=1, out_channels=32, kernel_size=3, stride=1, padding=1)output_2d = conv_2d(x_2d)print(f"2D Output Shape: {output_2d.shape}") # Shape: (1, 32, 64, 64)# 3D Convolution: Video or volumetric CT scan# Input shape: (Batch Size, Channels, Depth/Time, Height, Width)x_3d = torch.randn(1, 1, 16, 64, 64)conv_3d = nn.Conv3d(in_channels=1, out_channels=64, kernel_size=3, stride=1, padding=1)output_3d = conv_3d(x_3d)print(f"3D Output Shape: {output_3d.shape}") # Shape: (1, 64, 16, 64, 64) ``` Whether operating over a 1D audio sequence, a 2D digital photograph, or a 3D medical volume, convolution remains conceptually identical. It is a mathematical mechanism for filtering, localized pattern recognition, and dimensionality extraction across spatial and temporal dimensions. By shifting from manual, hand-engineered kernels to trainable weight matrices, computer science transitioned from traditional image processing to Modern Artificial Intelligence. [Convolution: Where Modern AI Begins](https://blog.stackademic.com/convolution-where-modern-ai-begins-512aa6cb9488) was originally published in [Stackademic](https://blog.stackademic.com) on Medium, where people are continuing the conversation by highlighting and responding to this story.