Hebbian learning can explain rhythmic neural entrainment to statistical regularities

Abstract

In many domains, learners need to extract recurring units from continuous sequences composed of discrete units. For example, fluent speech is perceived as a continuous signal (at least in unknown languages). Learners need to extract the underlying words from this continuous signal and then memorize them. One prominent candidate mechanism tracks how predictive syllables(or other items) are of one another. Syllables within the same word are more predictive of one another than syllables straddling word boundaries. Behaviorally, however, it is controversial whether such statistical learning abilities lead to memories of the underlying units beyond pairwise associations among syllables. The strongest evidence for this possibility comes from evoked electrophysiological responses time-locked to word boundaries (e.g., N400’s) and rhythmic neural activity with a periodicity of word durations. Here, I reproduce such results with a simple Hebbian network. When exposed to statistically structured syllable sequences, the network activation is rhythmic with a periodicity of the word duration and an activation maximum on word-final syllables. This is because word-final syllables receive more excitatory input from earlier syllables with which they are associated than syllables that occur earlier in words (and are less predictable). Hebbian learning can thus explain rhythmic neural activity in statistical learning tasks without any memory representations of words. N400’s time-locked to word boundaries might thus index the reduced predictability of word-initial syllables rather than word-onsets. Learners might thus need to rely on other cues to learn the words of their native language.