[From Bill Powers (961001.0500 MDT)]
Avery Andrews 091001 2:25 Eastern Oz Time --
I think that the `predictions' they're talking about are expectations
about the low-level details of the scene, based on its hi-level
categorization. E.g. it's supposta be a face so there must be a nose
somewhere. The details are quite beyond me, however!
The Kalman Filter method, which I understand only approximately, seems to
involve a process of extrapolation of an output from the present into the
future, given a description of an input and some of its time derivatives.
What you describe sounds more like a reasoning process to me, in which one
is presented with a perception of a whole, and from it deduces what some of
the parts must be, given previous experience with similar wholes. In a
hierarchical model of perception, of course, perception of the whole is a
function of the parts that are being perceived, and while imagination (an
internal world-model) can supply some non-essential parts, it can't
substitute for the initial perception of enough of the parts to create a
sense of the whole.
In the Rao article, mention is made of "feedback" and "feedforward"
connections between cortical layers, but there is no other assignment of
function to these connections. The possibility that the downgoing pathways
carry error or output signals from a higher system to a lower one is not, as
far as I can make out, considered. If Rao had stated explicitly that the
feedback and feedforward connections exist completely within parts of the
brain associated exclusively with perception rather than action, it would
have been clearer why he proposed this model of perceptual functions. As it
is, I can't even figure out which direction he considers to be "forward" and
which "back". Is the upgoing path the feedforward connection, or the
The only empirical phenomonenon that seems to be taken into account is the
existence of receptive fields -- the "blobs" I mentioned. It seems to me
that there are many ways to build a model that will exhibit the same
phenomenon. For example, why does he reject the perceptron approach -- or at
least not mention it? I think he's less interested in finding out how
perception works than in exploring the possibilities of a particular
computational method in which he has invested a lot of intellectual effort
that he doesn't want to waste. How's that for deducing controlled variables
on the basis of no evidence at all?