YM: Inhibition comes more naturally after Bidirectional Excitation, which helps motivate the need for Inhibition.
RO: its a chicken-egg thing -- can't actually run those models without inhibition in place already. could just run with "stealth" inhibition but seems better to forward-motivate. inhib.proj does a good job of this.
YM: I cut this, so we can stick to a single example to make the overview points: Mathematical concepts are a good example of particularly abstract and general categories that are clearly very powerful (e.g., the quantity "3" can be applied productively to a huge set of different inputs).
YM: I don't think either of the bidirection examples is very good. Figuring out where you know someone from is likely hippocampal, not about resolving ambiguity in processing face information. Explicitly scanning through the crowd isn't about higher-level knowledge per se. A better example might be trying to search for someone in a big crowd, and doing so by maintaining an image of what you are looking for (e.g., their red jacket), which helps to boost the relevant processing in lower-level stages.
and our ability to resolve ambiguity in inputs by bringing higher-level knowledge to bear on lower-level processing stages. For example, if we see a somewhat familiar face at the supermarket, but just can't place where we know them from, we can use top-down strategies to consider where we might know that person (it is often surprisingly difficult to recognize people out of context, as professors often discover when encountering former students off campus). Another real-world example is trying to search for a companion in a big crowd of people (e.g., at a sporting event or shopping mall) -- you have to explicitly scan through the crowd and focus your attention in different locations in order to pick them out.
RO: go for it!
YM: fig 33 seems obfuscatory
YM: To note the suggestions I just conveyed in person: In the next pass on this text, we should highlight the main take-home points throughout, since some of this text could otherwise come across as a bit rambly. One case in point is the Halle Berry neuron example, which I think should just be described as toward one end of a continuum of graded, distributed reps rather than suggesting across the text/captions that maybe there are Halle Berry neurons, maybe there aren't.
YM: Fig 21 should be reordered to match the text: feedback inhibition, then feedforward.
- add discussion of term "representation"
- add discussion of noise -- important for necker cube
- make clearer the connection between bidirectional and attractor -- why does attractor require bidirectional?
- revise connectivity fig again to include *3* areas and include the local input -> hidden connections.. maybe do as a separate fig?
The ability to entertain multiple such potential categories at the same time may be an individual difference variable associated with things like political and religious beliefs (todo: find citations).
A few brain-politics links (not related to categories though):