Inhibitory Oscillations and Retrieval Induced Forgetting

The connectionist perspective on cognitive psychology is appealing because it naturally explains many complicated phenomena with the relatively simple concept that activation spreads throughout a network of interconnected nodes, and that these interconnections are modified on the basis of repeated use. Furthermore, the theory is not merely verbal; connectionists have implemented these ideas in artificial neural network models of cognition, allowing for more strenuous, quantitative tests of hypotheses than afforded by models that are not mathematically implemented.

However, one challenge for this perspective would seem to be the concept of retrieval induced forgetting, in which retrieving some practiced word-pair (such as fruit-apple) impairs the later recall of another studied - but unpracticed - word pair (such as fruit-pear, even when given the cue "fruit-pe__"). Given the connectionist concept of spreading activation, one might expect the opposite result - that the interconnected networks representing fruits would enjoy a higher activation, and that therefore all the words paired with fruit would enjoy a facilitation in processing.

However, this is the opposite of what you see. Instead, these closely-related word pairs seem to be suppressed relative to baseline retrieval rates. According to some, this reflects an active inhibitory process (a concept which is distasteful for many connectionist modelers because it is both practically difficult and metabolically inefficient for such a representation scheme to be implemented neurally). One alternative theory is that the practiced word pair becomes selectively strengthened relative to all competing but unretrieved word pairs. To continue our example above, this theory suggests that the neural connections representing features unique to the practiced word pair (fruit-apple) are selectively strengthened, while the neural connections representing features shared by both word pairs (fruit-apple and fruit-pear) are selectively inhibited.

Although verbally elegant, it's not clear exactly how this process might be implemented neurally. However, a new paper in-press at Neural Computation by Norman, Newman, Detre, & Polyn provides one possible explanation for how this selective strengthening/weakening might be accomplished in connectionist terms. In their paper, "How Inhibitory Oscillations Can Train Neural Networks and Punish Competitors," the authors describe how biologically plausible oscillations between excitatory and inhibitory neural activity may serve to "tune" representations in the short term, such that "competitors" (fruit-pear) become less active while "targets" (fruit-apple) become more active.

Theoretical extensions to this framework invoke a characteristic response function of many cortical neurons to stimulus onset, in which excitatory activity quickly rises, but is then tamed by inhibitory activity. This inhibitory activity may overcompensate, and then back off, leading to an increase in excitatory activity again, before inhibitory activity clamps down once again. This process repeats, leading to a characteristic neural response function that clearly illustrates how a balance in activation emerges from competing oscillations of inhibitory and excitatory activity. This response function could theoretically account for the computation proposed by Norman et al. because of the way calcium is known to modify synaptic efficacy: at low but above baseline levels of calcium, synaptic efficacy is decreased, whereas at high levels synaptic efficacy is increased. According to this system, then, competitors would effectively be "inhibited" as a result of getting active only when the excitatory oscillation in the neural response function is overcompensating, because they would receive some calcium, but not enough to increase synaptic efficacy. Conversely, neural connections representing the features of target pairs would receive more calcium, leading ultimately to an increase in their synaptic efficacy. However, this theoretical extension has not yet been successfully implemented in connectionist models.

Related Posts:
Entangled Oscillations
Nature's Engineering
Models of Active Maintenance as Oscillation
Profile: Mark Tilden