Previous: Related Work
Up: Related Work Next: Basic Color Terms
Embodiment and Symbol Grounding
In [Lakoff 1987], conceptual embodiment is described as the idea
that the properties of certain categories are a consequence of the nature
of human biological capacities and of the experience of functioning in a
physical and social environment. It is contrasted with the ``objectivist''
idea that concepts exist independently of the bodily nature of any thinking
beings and independently of their experience (p. 12). Several sources from
the cognitive science literature at large are cited in [Lakoff 1987] as
having contributed to the development of this idea. My definition of
embodiment in the context of autonomous agents is more restricted, but more
precise (Section 
).
According to [Harnad 1990], symbolic representations must be
grounded bottom-up in non-symbolic representations of two kinds:
- iconic representations, which are ``analogs of the proximal
sensory projections of distal objects and events'', and
- categorical representations, which are ``learned and innate
feature-detectors that pick out invariant features of object and event
categories''.
Symbolic representations, according to Harnad, are grounded in these
two levels of elementary symbols. Symbolic representations consist of
symbol strings describing category membership relations, e.g., An X is
a Y that is Z. An example would be ``A zebra is a horse that is
striped''.
Part of my model of color perception and color naming
(Chapter ) is a color space defined by several
dimensions derived from or modeled after the activations of the cone
photoreceptors in the human retina. This is an analog (continuous)
transform of the sensory ``projection'' of the radiation reflected off the
``distal'' surface of the object being imaged; therefore, it qualifies as
an iconic representation. A second part of the model is a mapping from each
point in the color space to one or more pairs consisting of a color term
and a ``confidence'' or ``goodness'' measure. The codomain of this mapping
qualifies as a categorical representation, in Harnad's framework. My model
does not involve symbolic representations of the kind Harnad envisions, as
it is limited to the level of terms only, not sentences.
Harnad lists five tasks a cognitive theory has to explain: discrimination,
identification, manipulation, describing, and responding to descriptions of
objects, events, and states of affairs in the world. A CLR (Color Labeling
Robot) is in principle able to perform all these tasks:
- Discrimination: any two stimuli that result in different coordinates
in the model's color space are discriminable for the CLR
- Identification: the CLR is able to identify (categorize) any
stimulus as belonging to one (or more) of the categories represented in the
codomain of
, and give a confidence or ``goodness'' rating.
- Manipulation: an extended CLR (equipped with manipulators) that is
capable of acting in its environment, in addition to perceiving it, is able
to manipulate objects based on their perceived color properties.
- Describing: since the model only extends to the level of terms,
describing is limited to naming the colors the agent perceives in the
world, which the CLR is able to do.
- Responding to descriptions: the CLR responds to a description (color
name) by pointing out an example fitting the description (a colored object
in its environment).