Incident light reflected from a uniformly colored surface enters the eye and impinges on the retina, initiating a chain of neural activity. I assume that somewhere downstream (exactly where doesn't matter for my purposes) there will be what I call a neural activity pattern (NAP) that can be used (in principle) to distinguish among instances of incident light with different spectral composition. In "Empiricism and Philosophy of Mind", Chapter VI, Sellars introduces the concept of a sensory repeatable. In his terminology, a generic NAP (ie, one caused by incident light of unspecified spectral composition) is an instance of a determinable repeatable.
Through conditioning, a child acquires dispositions to respond to the occurrences of determinable repeatables. For example, in response to the presence of a surface that reflects light having a spectral power distribution (SPD) that spikes at long wavelengths a child may acquire the disposition to utter the word (color name) red. I hypothesize that the process of acquiring such a disposition involves the formation (via brain plasticity) of a neural structure - so to speak, a neural matched filter (see note below) - that responds to inputs that are in some sense "sufficiently close" to the determinable repeatable that occurred during the acquisition period by activating motor neurons that produce the appropriate utterance. (In that sense, Locke may have been on the right track with his impressions, only they are made in neuronal structures rather than wax tablets and include essentially "canned" responses.) Thus, there will be a family of determinable repeatables the occurrence of any member of which will result in utterance of the same color name. The conjunction of a family of determinable repeatables and a specific color name creates a family of determinate repeatables.
Note that this description says nothing about visual phenomenal experience, ie, mental imaging. I hypothesize that visual stimuli are processed along two different paths: one path is where relevant sensory-motor neuron connections are created and maintained, the other is where phenomenal experience is produced. There are instances of subjects who claim not to have any phenomenal experience on one side of their FOV but can correctly identify the direction of motion of an object there. This - and arguably Libet's and Nicolelis' experiments seem consistent with that hypothesis.
Based on the discussion in EPM chapter VI, it appears that some of the Locke-Berkeley-Hume issues raised there may have resulted from concerns arising from trying to account for the phenomenal experience of colors and the resulting intuition that somewhere there must be something that "is red". But in the view outlined above, "is red" is not a predicate which is true of any object but is merely an utterance in response to certain neural events. One might then say that to grasp the concept of a specific color is to master the use of a word (eg, red), and to grasp the concept of color in general is to master the use of a family of words (color names).
Now consider the "common descriptive content" of the three situations identified in EPM section 22, situations which are distinguished by the confidence levels implicit in the particular phrases describing them. A person looking at a surface in a controlled environment - standard lighting, no contrasting color surfaces, etc - might confidently say "that surface is red"; in an environment in which one or more of those conditions is not met might say less confidently "that surface looks red"; and in an environment in which there is reason to believe that illusion is possible (eg, the ghost scenes in "The Illusionist") might say even less confidently "It looks as if there is a red surface there." It is conceivable that the same color name (red) is used in all three cases because of metamerism, the ability of different SPDs to cause the same determinable repeatable. The fact that the complete utterances are different reflects that responsive dispositions are, in general, context-dependent.
Inverted spectra (the second "problem" in EPM sect 26 ) don't appear to affect anything in the view described. Although it's in principle conceivable (though possibly not consistent with the relevant neurophysiology) that the family of determinable repeatables that in most people are a component of their red determinate repeatable could be, for some individuals, the family of determinable repeatables that are, for most people, a component of some other determinate repeatable (ie, evoke utterance of a different color name). But the entities that are (in principle) constant when individuals learn to utter a color name are the incident SPDs, not the resulting determinable repeatables. Everyone will (in principle) respond appropriately to the same SPD.
A filter is a device designed to pass inputs at some frequencies and attenuate or reject inputs at other frequencies. Eg, a simple ideal bandpass filter passes inputs with frequency in some range (the band) unattenuated but rejects all other inputs. A matched filter is designed to produce the maximum response for a specific signal although it will respond to a lesser degree to other signals that are in a sense "close" to the signal to which the filter is matched.
One can think of the neural configuration that is produced in a learning process (eg, learning to say red in response to members of a family of SPDs) as "matched" to that family. SPDs that are not family members but are in some sense "close" to members of the family may cause a different learned response, perhaps "reddish", and in that sense resemble the red SPDs. SPDs that cause an unrelated response, say, violet, will not resemble the red SPDs.
Caveat: This idea is merely a heuristic; I have no reason to assume that anything like what I describe is actually constructed in the brain.