This circuit implements the model proposed in ((Whittington & Bogacz, 2017) [1].
Specifically, this model is supervised and can be used to process sensory
pattern (row) vector(s) x to predict target (row) vector(s) y. This class offers,
beyond settling and update routines, a prediction function by which ancestral
projection is carried out to efficiently provide label distribution or regression
vector outputs. Note that "FFM" denotes "feedforward mapping".
The GNCN-t1-FMM is graphically depicted by the following graph:
.. table::
:align: center
+---------------------------------------------------+
| .. image:: ../images/museum/gncn_t1_ffm.png |
| :scale: 75% |
| :align: center |
+---------------------------------------------------+
.. autoclass:: ngclearn.museum.gncn_t1_ffm.GNCN_t1_FFM
:noindex:
.. automethod:: predict
:noindex:
.. automethod:: settle
:noindex:
.. automethod:: calc_updates
:noindex:
.. automethod:: clear
:noindex:
References:
[1] Whittington, James CR, and Rafal Bogacz. "An approximation of the error
backpropagation algorithm in a predictive coding network with local hebbian
synaptic plasticity." Neural computation 29.5 (2017): 1229-1262.