Composing Linear Layers from Irreducibles

Travis Pence (University of Wisconsin-Madison) · Daisuke Yamada (University of Wisconsin - Madison) · Vikas Singh (UW-Madison)
algebraic perspectivebivectorsblock-hadamardclifford algebracompositional structuredeep modelsdifferentiable algorithmgeometric objectsgeometric primitiveshigher-level functionslinear layerslinear transformationsllm attentionlow-rank approximationsparameter efficiencyrotors

Contemporary large models often exhibit behaviors suggesting the presence of low-level primitives that compose into modules with richer functionality, but these fundamental building blocks remain poorly understood. We investigate this compositional structure in linear layers by asking: \textit{can we identify/synthesize linear transformations from a minimal set of geometric primitives?} Using Clifford algebra, we show that linear layers can be expressed as compositions of bivectors