On the Biological Plausibility of Artificial Metaplasticity - Foundations on Natural and Artificial Computation - page 125

separated by a LTP threshold. In [25] Abraham and Bear consider it as a ho-
mosynaptic property (i.e. involving only the synapse under study, without the
need of considering the influence of nearby synapses). Whereas the model by
Bienenstock, Cooper and Munro [21] yields a curve that is closer to reality, with
LTP threshold determined by postsynaptic activation and without LTD thresh-
old. In the BCM model, the LTP threshold is the same for all neuron synapses,
so that metaplasticity would affect even non-active synapses (heterosynaptic
plasticity). But despite its unquestionable biological characteristics, the BCM
model, cannot be regarded to be the ultimate model of synaptic metaplasticity.
According to Mockett and colleagues [27] metaplasticity is inherently a homosy-
naptic phenomenon in contrast to the heterosynaptic nature of the BCM rule.
Finally, Artola, Bröcher and Singer's [28]) extended model (ABS model) is not
analytical, as those just discussed, but is based on empirical experimental data.
In the ABS model, LTP and LTD thresholds shift to lower values for higher
levels of the activation of neighbouring synapses.
In its implementation characteristics, the proposed AMP model follows closer
to the BCM model. We do not pretend to determine its superiority, but neurobi-
ology inspires computer science and vice versa, and we report that the empirical
results of AMP show a great potential, in terms of and there-

fore performance in most cases, no matter in what multidisciplinary application
is applied [1, 17, 18, 28].

improving learning

7Con lu on

We describe and discuss the biological plausibility of an artificial model of meta-
plasticity, a relevant property of neurons. Tested on different multidisciplinary
applications, it achieves a more ecient training and improves Artificial Neu-
ral Network Performance. The model follows the BCM heterosynaptic biological
model. During the training phase, the artificial Artificial Metaplasticity Multi-
layer Perceptron could be considered a new probabilistic version of the presynap-
tic rule, as it assigns higher values for updating the weights in the less probable
activations than in the ones with higher probability.

References

1. Andina, D., Alvarez-Vellisco, A., Jevtic, A., Fombellida, J.: Artificial metaplasticity
can improve artificial neural network learning. In: Intelligent Automation and Soft
Computing, SI on Signal Processing and Soft Computing, vol. 15, pp. 683-696
(2009), ISSN: 1079-8587
2. Abraham, W.C.: Activity-dependent regulation of synaptic plasticity (metaplas-
ticity) in the hippocampus. In: Kato, N. (ed.) The Hippocampus: Functions and
Clinical Relevance, pp. 15-26. Elsevier, Amsterdam (1996)
3. Abraham, W.C., Bear, M.F.: Metaplasticity: The plasticity of synaptic plasticity.
Trends in Neurosciences 19, 126-130 (1996), doi:10.1016/S0166-2236(96)80018-X