Vision-Specific Concepts
Images are structured grids: nearby pixels are related, objects span multiple scales, and appearance can change without changing meaning. Vision models use these properties to extract, aggregate, and generalize visual evidence.
Start with kernels and convolution, then follow how receptive fields and pooling reshape spatial information. Attention adds content-dependent context, while data augmentation defines which visual changes a model should learn to ignore.
- AttentionDynamic feature weighting mechanism
- ConvolutionSpatial feature extraction operation
- Data AugmentationData Augmentation - Expanding training data through label-preserving transformations
- KernelLearnable filter in convolution layers
- PoolingSpatial downsampling operation
- Receptive FieldInput region influencing a neuron