The dictionary-learning lineage
None of the ideas behind this paper's core method are new on their own. What's new is scaling a decades-old technique up to a real language model and following through on an unpublished hunch someone else had left unfinished.
This theme traces that lineage back through sparse coding's roots in modeling the visual cortex, to an early attempt at applying it across a whole transformer's layers at once, to the unpublished report that first proposed training it specifically to fight superposition, and to a separate mathematical result invoked to explain why some directions turn out easier to interpret than others.
Cunningham et al. (2023) present their method as a scaling-up of existing ideas, not a new invention. Sparse dictionary learning itself is decades old, developed by Olshausen and Field (1997) and Lee et al. (2006) to model sparse coding in the visual cortex, expressing data as sparse linear combinations of an overcomplete learned basis. Two direct predecessors bring it to language models: Yun et al. (2021) apply sparse dictionary learning to residual-stream activations across all layers of a transformer simultaneously, and Sharkey et al.'s (2023) unpublished interim report first proposes training sparse autoencoders on language-model activations specifically to address superposition, credited as this paper's direct motivation. A separate theoretical result, Qu et al.'s (2019) reformulation of sparse overcomplete dictionary search as a search for directions maximizing the L4 norm, is invoked to explain why interpretability correlates with the heavy-tailedness of feature activations and why ICA, which optimizes non-Gaussianity directly, is the strongest non-dictionary-learning baseline.