Different histories of categorisation generate different feature
spaces to encode similarities and contrasts between objects.
The different categorisation histories in this experiemnt led to
different concept learning mechanisms. This would seem to be a
demonstration of the flexibility involved in feature detection and
categorisation. So, it seems that the type of features you are
exposed to has an influence as to whether or not your encoding and
recognition of a novel item is based on flexible or fixed space
The premise that features are created to subserve categorisation
applies to the creation of functional features but is neutral as to
their perceptual realisations.... In short, there are many possible
realisations of a functional feature.
Yes, it's true that the object property 'square' can be featurally
represented in many ways e.g. in terms of lines, angles, corners etc.
But whatever that representation might be it always has some
relationship to its perceptual realisation - no matter how tenuous
that link may be.
One particular problem that must be addressed is the degree to which
these functional features are themselves based on a (more) primitive
set of features.
You can make estimations and approximations as to the extent to which
they are based on primitve features, but I'm not sure how this could
be measured scientifically.
We will accordingly argue that functional features are not always
constructed out of a fixed catalogue of primitive features.
But what about the blobs argument? It is logical to assume that no
person can have infinite knowledge about all the possible shapes of
blob that could feasibly exist. But it is still valid to argue
the possibility that perhaps the features of blobs are just
recombinations of known curves (fixed primitive features) and
integration of these facets allow for feature recognition and
categorisation. Wouldn't this process of integration of recombined
primitives be a valid example of the mechanism by which functional
features can always be constructed out of a fixed catalogue of
It is not logically feasible to extract relevant categorisation
features from pixel-based (or similarly unstructured)
representations of the input.
This makes sense. Unstructured primitives would, in reality, be too
small, diverse and numerous to make it an economical mechanism of
feature recognition. Proper integration could never be achieved.
Any large-scale highly structured set of primitives is bound to be
too coarse to detect (and internally represent) all of the
distinctions that might be required by different categories of
Structure primitives are not flexible enough to allow for all
available feature possibilities. Their coarseness means that many of
the subtleties in features would go unrepresented. Surely there must
be some middle ground between the structured and unstructured
primitives that would allow for the appropriate construction of
This archive was generated by hypermail 2b30 : Tue Feb 13 2001 - 16:23:21 GMT