How language processing might go awry

Re: April Benasich’s ongoing studies of Specific Language Impairment in children ( If we believe Maass, et al. (2004, “On the computational power of circuits of spiking neurons”)  with respect to the ability of a (plausible, cortical) recurrent neural network to retain in its state a usable record of (order of) the last three seconds of its inputs, we may get some insight into possible underlying mechanisms for what Benasich hypothesizes to be  “the primary deficit in … SLI … the inability to integrate and discriminate two or more sensory events which enter the nervous system in rapid succession.” 


Maass’s work suggests that in normal processing, asynchronous discriminations can be replaced by synchronous discriminations.  I interpret this to mean that as soon as the neural circuitry has enough of the stimulus to recognize, it can be pulsed out synchronously to the next stage of processing.  Looking at psychophysical results like those of Richard Warren (1970, “Perceptual Restoration of Missing Speech Sounds.”  Science 167: 392-393) could be interpreted as indicating that a later pulse can refine or override (confirm or disconfirm) the data delivered in an earlier pulse. 


So here’s what I think happens.  Suppose Stimulus A is followed in quick succession by Stimulus B.  Sometime during or after Stimulus A (because the neuro-glial circuitry doesn’t have to wait until what we think is the “end” of a stimulus before reacting) the circuitry approaches or settles into a provisional identification of an Event Pattern of, call it, Type X.  Event Pattern Type X is the pattern of activation (state) created by A in the context of mutter-mutter (unknown or irrelevant—whatever the state was when A arrived).  This provisional identification state somehow gets sent on as a synchronous pulse.  That pulse “runs out of oomph” somewhere and leaves neuro-glial elements in its trail primed with EPSPs and IPSPs.  (All I intend by that is to indicate that the pulse isn’t sufficient in and of itself to cause recognition of A tout court (as Jerry Fodor would say).


In normal processing, sometime during or after Stimulus B (which follows Stimulus A in rapid succession) the circuitry will  approach or settle into a provisional identification of an Event Pattern of Type Y (which is the state created by B in the context of immediately preceding A and whatever came before).  That information gets sent on in a pulse.  In between, there may be (there probably are) confirming or disconfirming pulses (possibly at regular intervals).  The net result is (insert hand waving here) recognition of A followed by B.


So what could be happening in SLI?  Possibilities are:


1)       Stimulus history traces decay so rapidly that at the time Stimulus B arrives, it has insufficient context information and gets sent on as an Event Pattern of Type Z (B in the context of mutter-mutter).  In later processing, this acts as disconfirmation of Pattern Type X (A in the context of mutter-mutter) rather than confirmation of the temporal evolution from A to B.  So information about A is lost.  I suppose it’s also possible that the apparent disconfirmation could be treated as spurious, so information about B could be lost.  Or the conflict could lead to no decision and loss of the distinctness of A and B.  Checkmate in all cases.


2)       State information isn’t being read out on a rapid enough schedule, so what comes through is only Event Pattern Type V (in the context of mutter-mutter, A followed by dim intimations of B) or Event Pattern Type W (dim intimations of A followed by B followed by whatever).  In either case, one of the stimuli is represented only by dim intimations that don’t reach above whatever threshold is necessary to affect subsequent stages, so information about A or B is lost.


3)       There is a timing mismatch in later processing so that differential decay rates of EPSPs and IPSPs cause information from what should have been distinct pulses to get muddled and the pattern looks like Pattern Type U (A overlaid with B in the context of mutter-mutter overlaid with A).  So the distinctness of A and B is lost.


4)       The state attractor that has developed in the neural circuit that gets first crack at A followed by B classifies them both the same way (like /p/ and /b/, I suppose Merzenich might say).

Leave a Reply

You must be logged in to post a comment.