erwin at trwacs.fp.trw.com
Thu Dec 3 08:43:36 EST 1992
>Could you please elaborate on your reference to Desimone? I don't catch
>what you're referring to here. Also, I'm having difficulty following
>this on account of some of the terms that you're using. "...not just
>their conclusions, but also the evidence for those conclusions." What
>do you mean by "evidence" here? By "conclusions" do you simply refer
>to the spatial pattern of output activity across the olfactory bulb,
>or the waveforms of these output signals, or both? Then how is the
"evidence" distinct from the "conclusions"?
Actually, it was Barry Richmond's paper, rather than Desimone's that
addressed this (they presented their papers back-to-back at Pribram's
conference, and I didn't check my notes in writing this up). Barry
examined the data content of the pulse train generated by your typical
pyramidal cell and demonstrated that the pulse train could be separated
into three components: a carrier wave, the information produced by the
individual neuron, and the information passed through by the individual
neuron from its sources. The carrier wave, BTW, appears to be used for
correlating multiple reports on the same source object across the cortex.
The PC appears to report the spacial pattern of OB activation, but it also
reports the invariant pattern of sensory cell activations that generates
the OB activation.
>I'm lost concerning this "synthetic sensory data." I catch your drift
>that there's a top-down feedback influence from the AON to the OB; what I
>don't follow is how exactly this allows the system to then discard learned
>patterns. Are you saying that the learning at the OB is transient, but
>that there is permanent retention in the AOR? The patterns are being
>learned, and stored, somewhere here, right?
The AON has the capability of replicating a given pattern of sensory cell
activations while simultaneously putting the OB in a neutral state, ready
to learn that pattern. In other words, it can present synthetically
generated sensory data to the OB to train it on that pattern. Meanwhile,
the PC receives the OB pattern (which is non-invariant) and learns the
current correlation between OB pattern and the pattern of presented
sensory data. This is not a permanent correlation, because the OB forgets
over the long term (probably so it can reuse the neural resources), but it
holds for the short term. The permanent retention is probably downstream
of the AON and PC, although the AON needs to learn the correlation between
its transmissions to the OB and the reported sensory cell activations back
via the PC pathway so that it can train the OB on command. The PC needs to
retain the correlations between OB patterns and the patterns the cortex
expects for the medium to short term and it needs to do some other
translations on a permanent basis to allow it to control the AON based on
cortical commands. The PC-OB interface appears to have two purposes:
setting the OB to neutral for each breath and generating the resulting
chaotic process. The AON appears to have two functions: training and
assisting in generating the chaotic process.
Probably key here is that the OB, AON, and PC are all content-addressible
memory systems that seem to be following a competitive learning model.
Each component does a translation, and as the definition of patterns shift
over time, each updates its internal translations.
If it sounds like I'm a software systems engineer with interests in neural
networks and non-linear dynamics, it's not accidental. Unlike Karl
Pribram, who wants to known how the brain works, I want to build a working
Internet: erwin at trwacs.fp.trw.com
More information about the Neur-sci