Supporting text.

<p>Supplementary information file including additional analyses on behavioural effects without feedback and late onset of low coherence trials, effects of regularisation type on network behaviour, the hidden layer model, weight and gate differences between L1- and L2-regularised networks as well as Gaussian noise differences at weights and gates between insight and no-insight networks. This file includes Fig A (Switch-aligned performance and switch point distributions for L1- and L2-regularised neural networks with a 48 unit hidden layer each), Fig B (Illustrations of models and respective parameters), Fig C (Performance on highest motion noise trials and model predictions for every human participant), Fig D (Performance on highest motion noise trials and model predictions for every network), Fig E (Switch-aligned performance for insight group and no-insight group), Fig F (Switch-aligned performance and overlap between classification and self-reported colour use), Fig G (Trial-wise insight-like strategy improvements for 5% motion coherence trials), Fig H (L2 networks: Task performance and insight-like strategy switches) and Fig J (Comparison of gate weight magnitude and influence of λ on insight-like behaviour across L1- and L2-regularised networks). Figure legends see inside S1 Text.</p> <p>(PDF)</p>