WITHDRAWN: Masking Drives Shared Representations in Predictive Networks

Abstract When must a neural network develop shared representations for self-generated and externally-observed information? We investigate this question using small recurrent networks trained on masked sensorimotor prediction. We find that cross-modal masking during training is the primary driver of cross-modal shared representation alignment (mean effect + 0.10 on Pearson correlation), while further constraint tightening (lower temporal persistence, higher masking probability) adds negligible mean improvement (effect = − 0.001). The real contribution of tight constraints appears at large capacity: modest floor elevation and variance compression across seeds. We report a methodological correction: earlier descriptions of a dramatic “optional-to-compulsory” regime transition were confounded by a metric change (Linear CKA to Pearson correlation) between experimental conditions. With consistent metrics, the transition is gradual, not phase-like. These findings suggest that cross-modal shared representations emerge from cross-modal predictive pressure, with capacity constraints modulating the strength and consistency of shared representation expression.