Many persistent limitations of neural ML systems appear linked to a lack of constraint on internal dynamical organisation. Existing regularisation methods largely target input-output behaviour or impose local smoothness and stability. My proposal takes a different approach by explicitly shaping the degree of coupling between internal states to promote more robust and coherent learned dynamics in recurrent and continuous-time models. I introduce an inductive bias, inspired by Integrated Information but grounded in classical control theory, that penalises internal dynamics that are easily decomposed into weakly interacting subsystems. This is implemented using Gramian-based measures of intrinsic state coupling, computed via local linearisation of the system Jacobian. The result is a differentiable scalar that can be incorporated into standard training objectives at polynomial cost. The attached proposal includes mathematical derivations, practical extensions addressing scalability and stability, experimental protocols, and an assessment of limitations and open questions.
David Cressy (Thu,) studied this question.