NEURAL NETWORKS,
A MODEL FOR THE BRAIN?
It must be clear that although neural networks are
inspired on the functioning of our brain, people who
create these networks are well aware that they are not
an accurate model of our brain. Nevertheless, these networks and our
brain have in common:
- massive parallelism
- distribution of computation on many processing units, which have
multiple purposes (one network can perform various
tasks)
- no division between processing units and memory containing
knowledge
- the processing units are massively interconnected
- flexibility and unsensivity to malfunctioning of parts
- capacity of association of patterns related, classification and
generalization of them
- they contain implicit representations of the relations between
patterns, distributed on multiple units and connections
Ofcourse, a list of characteristics that they don't have in common
with our brain, would be much longer, but, nevertheless, they come
one step closer to being a model for our brain than the traditional
artificial intelligence paradigm which does not possess the
characteristics mentioned above. The classic AI works with explicit
procedures to imitate human reasoning, and they focus on more
sophisticated aspects of it. Until now it has delivered poor
results. Neural networks focus more on low
level functions of our reasoning, and, although it's early to make
judgements, it seems a more fertile research area. The general idea
is that higher levels of our reasoning are based on low level
aspects of it, and therefore, if we start simulating these low level aspects,
eventually we will arrive to a theory which explains our reasoning
in general.
Subsymbolism
Subsymbolism means that concepts, instead of being represented with
a symbol (number, letter, word, phrase, memory adress, etc.), are
represented with quantitive properties of, in this case, a network. The single
quantitive property is just a sub-part of the representation of the
concept, a subsymbol. Traditional AI assumes that our reasoning
consists in handling symbols representing concepts, and identifies
these symbols with our natural language. Subsymbolism is a valid
alternative to this assumption. Concerning neural networks, the
subsymbolism paradigm stresses the importance of distributing
representations on multiple units. This means for example that, if
a concept has a certain property, this property is not represented
by the activation of one unit, but by the activation of multiple units.
Logic or intuitivity?
Most people assume that our reasoning is mainly guided by deciding
based on some kind of logic and planning our actions in a conscious
way. But is it really like that? How many activities do we perform
without any conscious deciding and planning involved? Obvious
examples are walking, avoiding obstacles that we find on our path
and observing our everyday reality with our eyes and ears. Maybe we
could even say that most of our activities are guided by unconscious
processes in the brain. Then, if our goal is simulating human
intelligence, wouldn't it be more smart to start simulating
unconscious processes of our brain, and then simulate a conscious
process of extraction of relations between concepts from these
unconscious processes? I think that neural networks can play an
important role in bringing us closer to this goal.
[back]
[forward]
[mail me]
Thomas Riga, University of Genoa, Italy