Nevertheless, such coding schemes are biologically unrealistic. In the distributed model, neurons would be fully interconnected (i.e., there would be
no anatomical specificity of feedforward, feedback, or lateral connections).
There would be no topographic representation of information. Neurons
would fail to show stimulus specificity. Such anatomical and physiological
patterns of organization, however, are not observed in nervous systems,
which exhibit precision of connections between and within different classes
of neurons, with the spatial location of a neuron typically representing
aspects of the neuron’s specificity for parameters that are encoded. Thus,
the extreme forms of the single cell and distributed-representation
hypotheses are both rejected because they fail to reflect known biological
reality.
The synthesis of the single-cell and distributed-coding hypotheses arises
from an appreciation of the different end states of a behavioral continuum.
Perceptions of arbitrary objects are unlikely to be processed by highly
object-specific neurons. This conclusion derives from constraints on proximate mechanisms as described above and constraints on ultimate mechanisms, especially the unpredictable behavioral significance of an arbitrary
object. For arbitrary objects, a distributed representation is likely, where the
activation of ensembles of less specialized neurons is ultimately related
directly to the perceptual event. In contrast, a predictable environment may
allow for either genetic fixation or learning during ontogeny (or both) to
establish specialized processing for reliable objects. There is considerable
evidence for the existence of such specialized hierarchical streams of processing (see below) with mnemonic cells that exist at higher levels of these
hierarchies. Rather than yielding a combinatorial explosion, extension of
the hierarchical organization within the constraints provided by a predictable environment may increase coding efficiency. Perception is still ultimately related to the activation of ensembles of neurons, but the size of the
ensemble may be different from that in the case of an arbitrary stimulus,
and the neurons that encode the predictable stimulus may be more highly
specialized. Thus, there is a balance among the various computational
requirements associated with different behavioral requirements (Rolls
1992). Assessing neuronal variation across such behavioral variation is at
the heart of the neuroethological approach.
In the context of the proposed framework, neurons act not as single-cell
indicators of complete percepts representing components of signals that in
combination result in perception of objects (cf. Barlow 1972), nor as featureless cogs in a vast distributed machine, but as localized feature detectors. In this proposal, the mechanisms for combining features may be single
cells, population dynamics, or both. For more arbitrary objects, the features
might be described by the statistics of natural scenes (Simoncelli and
Olshausen 2001). Each species also lives in its own unique perceptual environment, and this serves to delimit the more complex features and combination of features that are represented at the level of single cells.
332
T.Q. Gentner and D. Margoliash
no anatomical specificity of feedforward, feedback, or lateral connections).
There would be no topographic representation of information. Neurons
would fail to show stimulus specificity. Such anatomical and physiological
patterns of organization, however, are not observed in nervous systems,
which exhibit precision of connections between and within different classes
of neurons, with the spatial location of a neuron typically representing
aspects of the neuron’s specificity for parameters that are encoded. Thus,
the extreme forms of the single cell and distributed-representation
hypotheses are both rejected because they fail to reflect known biological
reality.
The synthesis of the single-cell and distributed-coding hypotheses arises
from an appreciation of the different end states of a behavioral continuum.
Perceptions of arbitrary objects are unlikely to be processed by highly
object-specific neurons. This conclusion derives from constraints on proximate mechanisms as described above and constraints on ultimate mechanisms, especially the unpredictable behavioral significance of an arbitrary
object. For arbitrary objects, a distributed representation is likely, where the
activation of ensembles of less specialized neurons is ultimately related
directly to the perceptual event. In contrast, a predictable environment may
allow for either genetic fixation or learning during ontogeny (or both) to
establish specialized processing for reliable objects. There is considerable
evidence for the existence of such specialized hierarchical streams of processing (see below) with mnemonic cells that exist at higher levels of these
hierarchies. Rather than yielding a combinatorial explosion, extension of
the hierarchical organization within the constraints provided by a predictable environment may increase coding efficiency. Perception is still ultimately related to the activation of ensembles of neurons, but the size of the
ensemble may be different from that in the case of an arbitrary stimulus,
and the neurons that encode the predictable stimulus may be more highly
specialized. Thus, there is a balance among the various computational
requirements associated with different behavioral requirements (Rolls
1992). Assessing neuronal variation across such behavioral variation is at
the heart of the neuroethological approach.
In the context of the proposed framework, neurons act not as single-cell
indicators of complete percepts representing components of signals that in
combination result in perception of objects (cf. Barlow 1972), nor as featureless cogs in a vast distributed machine, but as localized feature detectors. In this proposal, the mechanisms for combining features may be single
cells, population dynamics, or both. For more arbitrary objects, the features
might be described by the statistics of natural scenes (Simoncelli and
Olshausen 2001). Each species also lives in its own unique perceptual environment, and this serves to delimit the more complex features and combination of features that are represented at the level of single cells.
332
T.Q. Gentner and D. Margoliash