Convolutional neural network

Problem type

Supervised image classification, detection, segmentation; also audio spectrograms and some sequences.

Input

Grid-structured tensors, e.g. images $[B, C, H, W]$.

Output

Class logits, bounding boxes, or per-pixel labels depending on the head.

Mental model

Slide small, shared pattern detectors across the image; stack layers so edges become textures, textures become parts, parts become objects.

Core objective

Minimize a task loss (cross-entropy for classification) over convolution, pooling, normalization, and dense layers.

Training process

Mini-batch SGD or AdamW with heavy data augmentation; usually fine-tuned from an ImageNet-pretrained backbone.

Preprocessing

Resize and normalize with the pretrained model's statistics; label-preserving augmentation (crops, flips, color jitter).

Assumptions

Local patterns matter and can appear anywhere (translation equivariance).

Key hyperparameters

Architecture (depth, width, kernel sizes), learning rate and schedule, augmentation strength, weight decay, input resolution.

Good use cases

Image classification and detection, especially with transfer learning; efficient on-device vision.

Poor use cases

Tabular data; long-range dependencies without additional mechanisms.

Strengths

Data-efficient inductive bias; parameter sharing; mature pretrained backbones; efficient inference.

Weaknesses

Needs labeled images; can learn shortcuts (backgrounds, watermarks); sensitive to distribution shift.

Computational cost

Per layer about $H_{\text{out}}W_{\text{out}} \times C_{\text{out}} \times C_{\text{in}} k^2$ multiply-adds.

Evaluation metrics

Accuracy, top-5 accuracy, mAP (detection), IoU (segmentation); per-class and per-condition breakdowns.

Failure modes

Shortcut learning; label-changing augmentations; camera or lighting shift in production.

Minimal implementation

from torchvision.models import resnet18, ResNet18_Weights
import torch.nn as nn
model = resnet18(weights=ResNet18_Weights.DEFAULT)
model.fc = nn.Linear(model.fc.in_features, num_classes)   # fine-tune for your classes

Compared with neighbors

• Vision transformer: global attention, scales well with very large data.
• MLP: ignores spatial structure; far more parameters for images.

Learn more

Convolutional networks