Learning, Representations and Associative Memory
The representational hypothesis in brain research states
that sensations, percepts, thoughts and actions are somehow represented
by neural activity in the brain. The state of neural activity is influenced
by sensory input and by learned experiences of the past. There appear to
be quite different ways how the brain learns, i.e., organizes and reuses
past experience, some forms included and some not in common definitions of
human memory:
i) Representational learning is believed to form the internal
representations of stimuli. The result of representational learning are
mappings between sensory inputs and neural responses as, for instance, characterized
by receptive fields.
ii) Learning can extract regularities gathered during
previous experience that can be used for denoising of sensory input, for
instance, contour restoration based on enhanced excitatory connections between
nearby co-aligned edges (see illusory contours).
iii) Learning can lay down discrete memories, like memories
of episodes or abstract concepts such as "a tree" or "a car". A memory is
a (learned) mental entity that can be recalled in an almost binary fashion,
the recall is either made or not. Although the recall is discrete memories
can nevertheless can be multifaceted.
Human memory describes our ability to recall episodes
and to form and access mental concepts such as "a tree" or "a car" (type
(iii) learning). Its most intriguing property is that although the number
of episodes and concepts held in our memory is huge, every moment we can
selectively access just the most relevant (however, see the Stroop effect,...).
Such seemingly effortless selective access of memories is often called an
association. Although associations can be evoked by sensations, the underlying
neuronal processes must be quite different from a sensory restoration process
mentioned above (type (ii) learning). Introspectively, associations occur
more discontinuously (all-or-none character) than the perception of stimuli
and they are not rigidly coupled to the presence of a stimulus; they can
outlast the duration of a sensation or they can cease or shift without
the stimulus being changed. Furthermore, an association often involves aspects
extraneous to the modality of the stimulus. These observations suggest
that associations should activate neurons not exclusively involved in the
processing of the modality of the stimulus and that the dynamics of activation
should be to some degree decoupled of sensory input.
D. O. Hebb proposed a synaptic learning rule by which
experience can form distributed neural representations of mental concepts
by enhancing mutual excitation between the cells to be active. Such mutually
excitatory cliques of neurons (that represent mental concepts) Hebb called
cell assemblies. Cell assemblies show exactly the recall properties described
above. Mutual excitation between neurons within an assembly provides a discontinuous
transition to a collective active state in the assembly and the activity
can be persistent even after input withdrawal. There is growing indirect
experimental evidence for the existence of cell assemblies. However,
direct proofs are lacking because methods that record distributed patterns
of neural activity at high temporal resolution are still lacking at present.
To link the gap between experiments and the abstract theory of cell
assembly, methods of Computational Neuroscience can be applied. The implications
of the notion of cell assemblies can be studied more systematically in a
class of mathematical neural network models called neural
associative memory. These models mathematically describe the formation
and associative recall of cell assemblies in neural network architectures.