Recently, a scheme for object representation and classification that
is, in a sense, the conceptual opposite of WTA, has been put forward
under the name Chorus of Prototypes
[9]. This two-stage scheme employs vectors of
first-stage distances to a small number of reference objects to span
the second-stage representation space. In this manner, a
low-dimensional representation space for objects is built over a
high-dimensional space of primitive features. Recurring stable
patterns of primitive features, which are expected to correspond to
persistent objects, are represented explicitly, and constitute the
prototypes that span the object space. Each persistent prototype may
be represented by a set of detectors, implemented by RF-like
mechanisms tuned to a number of the object's views; these may be
constructed in a self-organizing fashion following exposure to the
object, as described in the Emergence of ... section. In distinction to
the persistent entities, rare or ephemeral patterns of primitive
features are represented implicitly, by the distributed activity they
induce in the prototype detectors. The latter must respond in a graded
rather than all-or-none fashion; the ratio between the actual and the
peak activity of a detector then may be interpreted as a correlate of
the distance between the current input and the preferred pattern for
that detector.
 |
 |
The relationship between Chorus and the issue of lateral connections
may be clarified by the following line of reasoning. The computational
basis of representation by the activities of prototypical-object
detectors is simultaneous response of several such detectors to any
given input (most of these will respond, if at all, at a fraction of
their maximal output level). As pointed out above, the response of a
detector thus constitutes a representation of the distance between the
current input and the optimal input for that detector, with the
distance being computed in a representation space common to all the
detectors. If this representation space is mapped onto the 2D surface
of the cortex, e.g. in the manner resembling the columnar functional
structure of area IT in the monkey, reported by Tanaka's group
[15,41,42,43], then
lateral spread of activation between columns may support the
computation of the representation-space distances. Moreover,
long-range cortico-cortical connections may endow the representation
space with additional structure, created by fiat as a result of
``random'' associations between items that are not necessarily close
neighbors in the original space.
In addition to learning by creating lateral associations, a system based on the Chorus principle can also learn by extending its repertoire of persistent features. A permanent record of the activation of a certain pattern of first-stage feature detectors may be created on a higher level, by assigning a higher-order detector unit to represent that pattern (this corresponds to the creation of a persistent representation instead of an ephemeral one). The resulting system derived from Chorus becomes essentially isomorphic to the NMR model described in the Emergence of ... section; an important distinction is that it aims at classification (making sense of multiple objects), rather than recognition (making sense of multiple views of the same object).
Figure 9: Chorus. To exploit the
ensemble similarity information, the mapping from the responses of
individual recognizers to face identity can be learned from
examples. Parallels between this scheme and multiple-view
representations, and the possible involvement of lateral
connections in its implementation, are discussed in
the Lateral Comparisons ... section.