Causal Direction from Convergence Time: Faster Training in the True Causal Direction

Researchers introduce Causal Computational Asymmetry (CCA), a new method for identifying causal relationships by training neural networks in both directions and determining causality based on which direction converges faster during optimization. The method achieved 26/30 correct causal identifications across synthetic benchmarks and is embedded in a broader Causal Compression Learning framework.

• →CCA identifies causal direction by comparing neural network convergence speeds when predicting Y from X versus X from Y.
• →The method operates in optimization-time space, distinguishing it from traditional statistical independence approaches.
• →Under additive noise models, the reverse causal direction has higher irreducible loss and requires more gradient steps to converge.
• →CCA achieved 26/30 correct identifications across six neural architectures in synthetic tests.
• →The approach is integrated into Causal Compression Learning framework combining graph structure learning and policy optimization.