Convolutional Neural Networks — Complete Guide

CNNs are the backbone of computer vision — they automatically learn spatial features from images.

How CNNs Work

Image → [Conv → ReLU → Pool] × N → Flatten → FC → Output

Conv Layer:
├─ Applies filters/kernels to detect features
├─ Edge detection, texture, shapes
├─ Local receptive field
└─ Parameter sharing (same filter everywhere)

Pooling Layer:
├─ Reduces spatial dimensions
├─ Max pooling: keep maximum value
├─ Reduces computation
└─ Adds translation invariance

Fully Connected Layer:
├─ Combines features
└─ Makes final prediction

Convolution Operation

Input (5×5):          Filter (3×3):        Output (3×3):
1 2 3 4 5            1 0 1                  4  3  4
6 7 8 9 0     *      0 1 0          =       2  3  2
1 2 3 4 5            1 0 1                   4  3  4
6 7 8 9 0
1 2 3 4 5

Output[i,j] = Σ Σ Input[i+k, j+l] × Filter[k, l]

Architecture

import torch.nn as nn

class CNN(nn.Module):
    def __init__(self):
        super().__init__()
        self.features = nn.Sequential(
            nn.Conv2d(3, 32, 3, padding=1),
            nn.ReLU(),
            nn.MaxPool2d(2),
            nn.Conv2d(32, 64, 3, padding=1),
            nn.ReLU(),
            nn.MaxPool2d(2),
            nn.Conv2d(64, 128, 3, padding=1),
            nn.ReLU(),
            nn.MaxPool2d(2)
        )
        self.classifier = nn.Sequential(
            nn.Flatten(),
            nn.Linear(128 * 4 * 4, 256),
            nn.ReLU(),
            nn.Dropout(0.5),
            nn.Linear(256, 10)
        )

    def forward(self, x):
        x = self.features(x)
        x = self.classifier(x)
        return x

Key Takeaways

1. CNNs automatically learn spatial features from images
2. Convolution detects local patterns (edges, textures)
3. Pooling reduces dimensions and adds invariance
4. Transfer learning with pre-trained models is standard practice
5. ResNet enables training very deep networks
6. Data augmentation is crucial for small datasets
7. CNNs achieve superhuman performance on image classification
8. Modern architectures: EfficientNet, ConvNeXt, Vision Transformer