local receptive fields

[Pankaj Mehra]

In response to Tom Dietterich <tgd at turing.CS.ORST.EDU>:
> I am confused by what appear to be two different usages of the term
> "local receptive fields" ....

and Scott.Fahlman at SEF1.SLISP.CS.CMU.EDU:
> the two usages of the term "local receptive field" are more or
> less the same idea, but different input encodings change the
> implementation.

I think that the question of locality starts well before the data are
given to the learning program. If we measure variables (dimensions) using
one sensor for each dimension, then transforming intervals (overlapping
or otherwise) of that dimension into local receptive fields is a matter
of representational convenience (different encodings, as Scott says).

Real receptive fields (i.e. units responding to limited regions of an abstract
feature space) perform oversampling of the input space, thus providing real
redundancy in observed data. There are multiple sensors for each dimension.

Artificial receptive fields (a good example is the BOXES representation used
in Chuck Anderson's thesis, U. Mass. Amherst, 1986) merely recode the
irreedundant data via a value-coded representation.

Artificial overlapping receptive fields merely oversample the data which is
collected non-redundantly.

In natural systems, real receptive fields can therefore alleviate sensor
errors (even dead cells) with a possible loss of resolution. Thus, they are
in some sense ``robust'' to noisy sensors.

Personally, I believe that multiple independent sensors should have
statistical significance as well, although I have not seen any discussion
of that in the literature I am aware of.

- Pankaj Mehra
University of Illinois