pyhgf.updates.vectorized.learning.vectorized_weight_update#

pyhgf.updates.vectorized.learning.vectorized_weight_update(parent_state, child_state, weights, coupling_fn, lr=None, parent_has_constant=False, adam_m=None, adam_v=None, adam_t=0, adam_beta1=0.9, adam_beta2=0.999, adam_epsilon=1e-08)[source]#

Unified weight update for vectorized layers.

Branches on the lr parameter:

  • Fixed (lr is a float): \(\Delta w = \text{lr} \cdot \text{PE} \cdot \pi_\text{child} \cdot g(\text{parent})\)

  • Dynamic (lr is None): Kalman-gain rule. \(K = \pi_\text{parent} / (\pi_\text{parent} + \pi_\text{child})\) \(\Delta w = K \cdot \text{PE} \cdot g(\text{parent})\)

When adam_m and adam_v are provided (not None), the fixed-LR gradient is filtered through Adam before being applied.

Parameters:

  • parent_state (LayerState) – Current state of the parent layer.

  • child_state (LayerState) – Current state of the child layer (with observations).

  • weights (Array) – Current weight matrix, shape (n_children, n_parents) or (n_children, n_parents + 1) when the parent layer includes a constant input node.

  • coupling_fn (Callable) – Coupling function applied to parent means.

  • lr (float | None) – Fixed learning rate. When None (default) the dynamic precision-weighted Kalman-gain rule is used instead.

  • parent_has_constant (bool) – If True, the parent layer has a constant input node. The parent mean is augmented with 1.0 and the precision with the parent precision mean so the bias column of weights is updated.

  • adam_m (Array | None) – First moment estimate for Adam, same shape as weights. Pass None to disable Adam.

  • adam_v (Array | None) – Second moment estimate for Adam, same shape as weights.

  • adam_t (int) – Global Adam timestep (already incremented for this step).

  • adam_beta1 (float) – Exponential decay rate for the first moment.

  • adam_beta2 (float) – Exponential decay rate for the second moment.

  • adam_epsilon (float) – Small constant for numerical stability.

Returns:

  • new_weights – Updated weight matrix.

  • new_adam_m – Updated first moment (or None when Adam is not used).

  • new_adam_v – Updated second moment (or None when Adam is not used).

Return type:

tuple[Array, Array | None, Array | None]