Understanding how complex machine learning systems make choices is a major challenge.
Interpretability research aims to make these decision processes transparent and safer.
Researchers analyzed three main approaches: feature attribution, data attribution, and mechanistic interpretability.
All methods struggle because the number of possible interactions grows exponentially as models scale.
The team introduced SPEX, which uses signal processing to find key interactions quickly.
SPEX relies on two observations: most interactions are sparse, and influential ones involve few features.
ProxySPEX adds a third observation about hierarchy in model structures.
ProxySPEX achieves similar results with roughly ten times fewer ablations than SPEX.
