In simple linear architectures like softmax regression, each input instance must be represented as a fixed-length 1D vector. When dealing with spatial data such as $28 imes 28$ pixel images, the 2D spatial structure is discarded and each image is flattened into a single vector of length $784$. This flattening step allows the data to be processed via standard matrix multiplication, although it ignores structural relationships that more advanced architectures like convolutional neural networks would otherwise exploit.