Hebbian Model Learning - Computational Explorations in Cognitive Neuroscience - page 114

(potentiation) and weakening (depression) of weights in
a nontransient (long term) manner. The form of LTP/D
found in the cortex has an associative or Hebbian form,
which depends on both presynaptic and postsynaptic
neural activity. Hebbian learning can be viewed as per-
forming model learning , where the objective is to de-
velop a good internal model of the important (statisti-
cal) structure of the environment. This type of learning
is often called self-organizing because it can be used
without any explicit feedback from the environment.
This chapter is the first of three chapters on learn-
ing mechanisms. It focuses on Hebbian model learning.
In the next chapter, we develop an alternative form of
learning called error-driven task learning, which does
a much better job at learning to solve tasks. The third
chapter then explores the combination of Hebbian and
error-driven learning, along with other more specialized
mechanisms necessary to address the important and dif-
ficult forms of task learning known as sequence and
temporally delayed learning.
This chapter begins with a simple Hebbian learn-
ing algorithm that captures basic aspects of biologi-
cal learning and picks up on correlational information
in the environment by performing a version of princi-
pal components analysis (PCA). This simple form of
Hebbian learning is then elaborated in an algorithm
that performs conditional principal components analy-
sis (CPCA), motivated by both computational and bi-
ological considerations. Finally, two correction fac-
tors are introduced that preserve the essential aspects
of Hebbian learning while making the algorithm more
effective.

for 1 second) will cause a long-lasting potentiation ef-
fect, with increases in synaptic efficacy of 50 to 100
percent or more. We now know a considerable amount
about the biological mechanisms underlying LTP, and
the related phenomenon of LTD, which is a long-lasting
depression (weakening) of synaptic efficacy.
The most common form of LTP/D in the cortex is
known as NMDA-mediated LTP/D .Thereisavery
nice connection between the way the NMDA recep-
tor works and the functional characteristics of this form
of LTP/D. These functional characteristics are generally
summarized by the term associative or Hebbian ,which
means that the activity states of both the presynaptic and
postsynaptic neurons are important for enabling poten-
tiation (or depression) to occur (i.e., there is an associ-
ation between these neurons). Hebb (1949) introduced
the idea that co-active representations should become
more strongly linked: if a given neuron consistently par-
ticipates in the firing of another neuron, then the con-
nection between the two neurons should be strength-
ened.
Focusing on the LTP case for the moment, the ob-
served associativity can be explained by the fact that
the NMDA receptor depends on both presynaptic and
postsynaptic activity before it will open. Presynaptic
activity is required because the NMDA channel will
not open unless the excitatory neurotransmitter gluta-
mate, released when the presynaptic neuron is active, is
bound to the receptor. Postsynaptic activity is required
because the postsynaptic membrane potential must be
sufficiently elevated (excited) to cause magnesium ions
( Mg + ) to move out of the opening of the NMDA recep-
tor channel, which they would otherwise block.
Once the NMDA channels do open, they allow cal-
cium ions (Ca ++ ) to enter the postsynaptic neuron. As
we noted in chapter 2, there is usually an extremely low
base concentration of calcium inside the neuron. Thus,
an influx of calcium can make a big difference, in this
case by triggering a complex cascade of chemical pro-
cesses, ultimately resulting in the modification of the
synaptic efficacy (weight) of the primary excitatory in-
put receptors, the AMPA receptors (figure 4.1). As dis-
cussed in section 2.3.2, there are a number of factors,
both presynaptic and postsynaptic, which can lead to the
modification of overall synaptic efficacy. The debate as

4.2 Biological Mechanisms of Learning

The modern era of the study of the biology of learning
at a cellular level began with the discovery of long-term
potentiation (LTP) (Bliss & Lomo, 1973). This term
contrasts with the many other transient forms of poten-
tiation that were known at the time, but that were not
suitable for the permanent storage of knowledge. Po-
tentiation refers to an increase in the measured depo-
larization or excitation delivered by a controlled stim-
ulus onto a receiving neuron. Bliss and Lomo discov-
ered that high-frequency stimulation (typically 100Hz