pyhgf.updates.vectorised.volatile.vectorised_layer_prediction_error#

pyhgf.updates.vectorised.volatile.vectorised_layer_prediction_error(layer, params, volatility_updates='eHGF', time_step=1.0, has_volatility_parent=True, max_posterior_precision=10000000000.0, mean_field_updates=False)[source]#

Compute prediction errors and apply the volatility-level posterior update.

This is the vectorised equivalent of pyhgf.updates.prediction_error.volatile.volatile_node_prediction_error(). It first computes value and volatility prediction errors, then dispatches to the appropriate volatility-level posterior update depending on volatility_updates.

Parent-count normalisation is applied in the prediction step (drift divided by n_parents), so the PE is the plain residual mean - expected_mean.

Parameters:

  • layer (LayerState) – Current layer with mean and expected_mean set.

  • params (LayerParams) – Layer parameters; provides the value level’s tonic volatility to the eHGF and unbounded volatility posteriors when one is allocated.

  • volatility_updates (str) – One of "eHGF" (default), "standard", or "unbounded".

  • time_step (float) – Current time step. Only required when volatility_updates="unbounded".

  • has_volatility_parent (bool) – If True (default), compute the volatility prediction error and apply the volatility-level posterior update (mean_vol, precision_vol). If False, only the value prediction error is computed and the volatility level is left unchanged.

  • max_posterior_precision (float) – Upper bound applied to the volatility-level posterior precision. Default 1e10.

  • mean_field_updates (bool) – Whether the prediction step ran the mean-field scheme; threaded into the eHGF and unbounded variance reconstructions so they subtract exactly what the prediction added.

Returns:

Updated layer state with prediction errors and volatility posterior.

Return type:

LayerState