Also, The Stability-Plasticity Dilemma
- is a name used to describe a problem encountered in neural network simulations. Many of these systems, once trained on a given set of exemplar responses, are simply not capable of learning anything new. This prevents the network from being able to continuously learn while it interacts with its surroundings.
The term is a bit of a misnomer, in that stability-plasticity merely highlights a problem (or dilemma) with conventional artificial neural network learning models. The general behavior of achieving stability and plasticity simultaneously in an adaptive system is not really a dilemma at all. The human brain is a perfect example of a system that quite handily achieves that goal. For that matter, so is the mouse brain. Since it involves asking the question: “How is simultaneous stability and plasticity facilitated within biological learning systems?” perhaps a better label might be, "The Stability-Plasticity Question."
click to enlarge
In other words the real crux of the question has been in how to design an artificial
system that—like the mouse brain—is simultaneously sensitive to, but not radically disrupted by, new learning.
This problem manifests in conventional ANN
s as catastrophic forgetting
. That is, the radical loss of most existing training when an attempt is made to add a single new item to the network's existing (i.e., pre-trained) response repertoire.
. . . . . . .
Enter: Multitemporal Synapses
A new learning method and mechanism, called “multitemporal synapses
” fully solves this problem. In fact, it is capable of providing a considerably greater range of stability-plasticity than is normally associated with human brains.
Like natural neural networks, this method permits the inclusion of both plasticity and stability as two separate, and distinct, components, which act independently of each other. In other words, they are not mutually exclusive. Because of this, multi-temporal synapses are able to continuously
(not merely continually
) learn and adapt, while the network-driven system interacts with its complex environment.
. . . . . . .
"Don't believe everything|
you read on the Internet"
Problem or Dilemma?
There is some disagreement on the Internet, as to whether we should call this a problem, or a dilemma. I recently got an e-mail from someone (May-2013) who insisted it should be a dilemma, but who would not share who he (or she?) was. His (her?) argument was that everybody agrees with him (or her, whoever he or she is).
There does seem to be some efferent
filtering going on within human perception that causes us to see mutually exclusive choices where no such mutual exclusion actually exists. Consider the rush, in the U.S., to choose between metric and imperial measuring systems back in the 1970
s, for example. Because of this added confusion, caused (presumably) by the very nature of how we think, it is probably best to begin with definitions of the terms used.
According to the Oxford-American Dictionary:
dilemma : noun : a situation in which a difficult choice has to be made between two or more alternatives, esp. equally undesirable ones
problem : noun : a matter or situation regarded as unwelcome or harmful and needing to be dealt with and overcome
So, for this to be a dilemma would mean that we are forced
to choose either
the other. That is, if this is a dilemma it is simply not possible
to choose both
. Or, at the very least, we would need to consider a tradeoff between stability and plasticity, where the more we have of either one, the greater would be the negative effect on the other.
While this may have been
the case for artificial
neural networks prior to multi-temporal synapses
, it certainly has not been the case for most biological nervous systems. Got that? Stability and plasticity are NOT
mutually exclusive. Nature shows us that we CAN
have both. The traditional limitation was purely an artificial one, caused by the limitations of our existing (artificial) neural network algorithms.
Now, however, with the advent of multi-temporal synapses
, the limitation has been eliminated for artificial neural networks as well. This problem
has been solved — Not by selecting between one or
the other (as we would have to do if it were an actual dilemma), but by getting rid of the limitation of traditional artificial
network structures. While traditional ANN
s forced us to make the trade-off, we can now include both—and as much of each as we want—in our neural network designs.
. . . . . . .