AssociativeFormer: Integrating Sparse Memory Attention and Versatile Tokenization for Hyperspectral Image Classification

Developing hyperspectral image classification models has long been hindered by problems regarding spectral redundancy and complexity, as well as the limitations of previously proposed transformer methods. We presentAssociativeFormer, an architecture designed to address hyperspectral analysis by merging memory association and efficient spatial-spectral processing. New features of the architecture include: Sparse memory attention with Top-K Routing that reduces computational complexity from quadratic to linear scaling while retaining its ability to discriminate effectively. The versatile convolution tokenization allows the architecture to capture multi-scale spatial-spectral features and filters from dynamically generated convolution tokens. The associative memory provided by a Hopfield network allows real time storage and retrieval of patterns, while the Gompertz Linear Unit (GoLU) activation decreases impediments on the gradient flow to create a more useful representation of the features. The network's global workspace design preserves inter-layer information for independent flow and spatial-spectral dependency tracking. TheAssociativeFormermodel delivers state-of-the-art accuracy with linear complexity and competitive inference speed, providing a superior accuracy–efficiency trade-off compared to existing hyperspectral classification methods.