The package 'edl' provides a set of functions that facilitate the evaluation, interpretation, and visualization of small error-driven learning simulations.
vignette("edl", package="edl") -
summarizes the core functions for training and visualization of results.
Also available online: https://jacolienvanrij.com/Rpackages/edl/.
Error-driven learning is based on the Widrow & Hoff (1960) learning
rule and the Rescorla-Wagner's learning
equations (Rescorla & Wagner, 1972), which are also at the core of
Naive Discrimination Learning (Baayen et al, 2011). Error-driven can
be used to explain bottom-up human learning
(Hoppe et al, under revision), but is also at the
core of artificial neural networks applications in the form of the
Delta rule.
This package provides a set of functions for building
small-scale simulations to investigate the dynamics of error-driven
learning and it's interaction with the structure of the input. For
modeling error-driven learning using the Rescorla-Wagner equations
the package 'ndl' (Baayen et al, 2011) is available on CRAN at
https://cran.r-project.org/package=ndl. However, the package
currently only allows tracing of a cue-outcome combination, rather
than returning the learned networks.
To fill this gap, we implemented a new package with
a few functions that facilitate inspection of the networks for small
error driven learning simulations. Note that our functions are not
optimized for training large data sets (no parallel processing), as
they are intended for small scale simulations and course examples.
(Consider the python implementation pyndl
https://pyndl.readthedocs.io/en/latest/ for that purpose.)
Doroth<U+00E9>e Hoppe, Petra Hendriks, Michael Ramscar, & Jacolien van Rij (2021): An exploration of error-driven learning in simple two-layer networks from a discriminative learning perspective. To appear in Behavior Research Methods.