On the Biological Plausibility of Artificial Metaplasticity - Foundations on Natural and Artificial Computation - page 121

w

w 1 w 2

w 3

LTD LTP

Postsynaptic Activity

w 3 > w 2 w 1

>

Fig. 1. Heterosynaptic BCM model [21] illustrates changes in synaptic strength due to
postsynaptic activity in biological neurons
.

2.1 Intrinsic Plasticity

A third concept about biological plasticity is the intrinsic plasticity, and it must
not be misunderstood with the previous ones. When neurons are receiving and
responding to synaptic inputs, the synaptic metaplasticity makes it dicult for
synaptic weights to become either null or saturated. But the metaplasticity prop-
erty cannot fully avoid these two extreme situations. For totally precluding the
possibility of either weight annihilation or saturation, another important homeo-
static property of real neurons should be taken into account: the so-called intrin-
sic plasticity [12]. Intrinsic plasticity regulates the position (rightward shift) of
neurons activation function, according to previous levels of activity [6]. Intrinsic
plasticity is not modelled in the AMMLP.

3 Metaplasticity and Shannon Information Theory

As is well-known within the ANN field, in 1949 Hebb postulated that during the
learning phase, synaptic connections between biological neurons are strength-
ened due to the correlated activity of presynaptic and postsynaptic neurons [11].
This plasticity property of synaptic connections is modeled in many ANNs as a
change in the connection weights of the artificial neurons or nodes. Therefore,
synaptic plasticity of biological neural networks has been simulated in artificial
networks by changing the weight values of the simulated neuronal connections.
These weights are the most relevant parameters in ANN learning and perfor-
mance. Modeling these new discovered properties of biological neurons that fol-
low metaplasticity rules provides a large potential for improving ANN learning.
In addition, the results of these simulations may also corroborate the biolog-
ical hypothesis of neuronal metaplasticity. Utilizing the potential of this new
modeling approach, artificial metaplasticity (AMP) models have been devised
and tested. A model that closely followed biological metaplasticity and intrinsic
plasticity was successfully tested in the reinforcing the calcium dysregulation
hypothesis for Alzheimer's disease [13]. However, of all AMP models tested by