A computational biologically-plausible model of working memory for serial order, repetition and binding

Current theories accurately view working memory as a multi -component structure, including a phonological store, a visuospatial sketchpad and a central executive. These components have been roughly identified with known brain areas and neuropsychological functions. However, the neural mechanisms for memory encoding and retrieval remain largely controversial, especially the mechanisms for serial order. These open questions warrant courageous attempts to tackle these problems computationally; that is, constructing large-scale biologically plausible computational models to explain detailed memory processes in these components, their interactions and the required control processes. The thesis is focused on constructing computational models to explain a few critical phenomena and mechanisms in working memory, observed in the Immediate Serial Recall task. The main constraining phenomena for the models are: the primacy effect, the recency effect, the phonological similarity effect in word lists but not in non-word lists, repetition coding by patterns, and position effects in transposition errors (e.g., an item is likely to be transposed with another item at the same position in another group). The modeling method is incremental: it first begins with a biologically plausible model for sequence encoding and retrieval based on the anatomy of the prefronto-basal ganglionic system. Phenomena that confirm this model include the primacy effect, the recency effect, similarity effects and the transposition gradient. This model is then extended with a specialized sequencing mechanism for phonological information, in order to explain the lack of phonoloigcal similarity effect in non-word lists. This Dual Representation model, to some extent, can also mitigate the difficulty of recalling repeated items, though it does not explain repetition encoding by patterns. The pattern encoding mechanism and the required binding process are discussed for a later model, which also uses binding to account for the positional effects, where position codes are considered to be temporarily bound to items. At last, it is discussed how the working memory models can be merged with a long-term memory component to transfer sequences stored in working memory into long-term memory. At last, limitations of the model are analyzed in the Implications and Discussion chapter. The analysis reveals a difficulty in the sequencing of items of different categories by lateral inhibition, because different categories of items are represented in different cortical areas. This result necessitates an additional subsystem that can encode category sequences, because the system needs to encode the category sequence to cue the retrieval of items. For example, a sequence like "Trees are plants" can be recalled better if it is accompanied by a category sequence "N V N". Due to the difficulty with repetitions of the basic cortico-basal ganglia sequencing mechanism, the category subsystem would work best if it is hierarchically organized, where each constituent does not contain repeated categories, or the repetition can be coded away by simple patterns. The category sequence then becomes a hierarchy, for instance, S(N VP(V N))). At this stage, the working memory theory meets linguistic theories. It implies, however, that the phrase-structure syntax should not itself be the ultimate criterion for linguistic acceptability or comprehensibility. If comprehensibility is defined as whether the brain can bind simple concepts into complex concepts following physical laws that govern synchronization, then, the brain should be able to make use of all available information from the sentence, context, long-term knowledge and the even the encoding mechanism. Therefore, an acceptable grammar can include phrase structure rules, and rules concerning the context, repetition patterns, position codes, and even syllabic counts. These implications of the working memory model deserve further exploration. Further analysis of the model also reveals a need for a specialized subsystem to encode the timing within sequences. It suggests that the cortico-cerebellar loop can be a good candidate for encoding temporal information, in addition to the serial order information encoded by the cortico-basal ganglionic system. A cerebellum-based mechanism for absolute and relative timing is discussed. This work will hopefully help further research about various sequence processing mechanisms of the brai