Trace a filter
Compute one output value from a 3x3 filter over a 5x5 grid, then shift the filter one column.
Remember CNN
Full name: Convolutional Neural Network
Memory aid: Convolutional filters Notice local Neighborhoods.
CNNs can be used beyond image classification.
Study area: AI & machine learning · Batch 4
Relevant study for: Azure AI Fundamentals AI-901 · Azure AI Apps and Agents Developer AI-103 · Azure AI Fundamentals (retired) AI-900 · Azure AI Engineer (retired) AI-102
Practice with exam flashcards →Data / Knowledge library
CNN (Convolutional Neural Network)
neural networks that learn local spatial features
Why learn CNN (Convolutional Neural Network)?
CNNs reuse learned filters across local regions, producing feature maps that can capture increasingly complex patterns. This structure is useful for image classification and other vision tasks. The Data category groups this concept with its implementation tools: PyTorch and TensorFlow, often used through Python. None of these specific tools is required by the mathematical architecture.
The problem it solves
Learns features from spatially structured inputs instead of relying entirely on manually designed image descriptors. Local connectivity and shared weights reduce parameters relative to a fully connected layer on the same input. Pooling or strided convolutions can reduce spatial resolution; pooling is not mandatory.
Where it came from
LeNet demonstrated convolutional networks for handwritten character recognition in the 1990s. AlexNet's 2012 ImageNet result helped popularize deeper CNNs. Later architectures explored greater depth, multi-scale processing, and residual connections. CNN describes a family of architectures, not a single product or release.
Build your understanding
Work through these exercises in order. Try the task before opening a reference.
- Apply a 3x3 filter to a small grid by hand and explain weight sharing.
- Calculate the output dimensions of a convolution with specified stride and padding.
- Build a small image classifier and inspect every layer's output shape.
- Compare training and validation curves and diagnose overfitting.
- Evaluate a pretrained CNN on held-out images and report per-class errors.
More ways to practice
Check tensor shapes
For a 32x32 RGB image, calculate the output of 16 filters of size 3x3, stride 1, no padding, and dilation 1. Verify with a framework.
Train a baseline
Train a small image classifier with separate training and validation data; log layer shapes and compare against a majority-class baseline.
Diagnose generalization
Compare learning curves before and after training-only augmentation; explain whether validation performance improves.
Audit an evaluation
Evaluate transfer learning on held-out data, grouping related images before splitting. Report a confusion matrix, failure examples, and inference latency.
Documentation and sources
- Stanford CS231n: CNN concepts and history (cs231n.github.io)
- TensorFlow: CNN image classifier (www.tensorflow.org)
- PyTorch: convolutional network tutorial (docs.pytorch.org)
Catalog research notes
- Reference snapshot: September 9, 2026. CNN is an architecture family with no single latest version.
- TensorFlow's introductory tutorial trains a convolutional image classifier; PyTorch's tutorial demonstrates a LeNet-style network and automatic differentiation.