D_RTRL

Contents

D_RTRL#

class braintrace.D_RTRL(model, name=None, vjp_method='single-step', fast_solve=True, normalize_matrix_spectrum=False, trace_dtype=None, **kwargs)[source]#

The Diagonal RTRL (D-RTRL) online gradient computation algorithm.

D_RTRL is the canonical name for the parameter-dimension eligibility trace algorithm implemented by ParamDimVjpAlgorithm. It computes the gradients of the weights with the diagonal approximation and the parameter dimension complexity, following the learning rule:

\[\begin{split} \begin{aligned} &\boldsymbol{\epsilon}^t \approx \mathbf{D}^t \boldsymbol{\epsilon}^{t-1}+\operatorname{diag}\left(\mathbf{D}_f^t\right) \otimes \mathbf{x}^t \\ & \nabla_{\boldsymbol{\theta}} \mathcal{L}=\sum_{t^{\prime} \in \mathcal{T}} \frac{\partial \mathcal{L}^{t^{\prime}}}{\partial \mathbf{h}^{t^{\prime}}} \circ \boldsymbol{\epsilon}^{t^{\prime}} \end{aligned} \end{split}\]

For more details, please see the D-RTRL algorithm presented in our manuscript.

This subclass inherits all behavior from ParamDimVjpAlgorithm without modification; it exists to provide the canonical D_RTRL name. See ParamDimVjpAlgorithm for the full parameter list and a usage example.

References

Contents