Models of Active Maintenance as Oscillation

Lisman and Idiart, Luck and Vogel, and Nelson Cowan have all suggested that working memory could be the result of the multiplexing of gamma oscillations (20-60 Hz) by theta oscillations (5-10 Hz) in the prefrontal cortex, such that capacity is determined by the number of gamma cycles that can occur within a single theta cycle.

Supporting this highly reductionistic claim are the observations that gamma oscillations are made more prominent by focused attention, that gamma oscillations are known to be important for transmitting information across large cortical distances and for visual binding of features into singular objects. Gamma synchrony is also known to increase performance in target detection as well as recall. Further, by playing auditory "clicks'" at near-gamma frequencies, it is possible to upwardly or downwardly entrain gamma rhythms and directly observe their effects on working memory span - exactly this was done by Burle and Bonnet.

Raffone and Wolters have presented a neural network model to simulate the limited capacity of visual short term memory in which oscillations occur between prefrontal cortex and inferotemporal (IT) cortex, and in which unrelated feature representations in IT mutually inhibit one another at a phase-lag that is automatically scaled to the number of features currently being encoded. In this view, capacity limits still result from the multiplexing of a fast rhythm with a slower one, but the fast rhythm is the overall firing rate while the slower rhythm is reflective of the phase synchronization of various feature encodings.

Usher and Cohen have proposed an alternate computational neurobiological model of visual short term memory capacity limits. According to their framework, the limitation occurs not because of precise frequency interactions but more generally as a result of flexible lateral inhibition. Normally, lateral inhibition is modeled as a k-winners-take-all system, in which a layer can support k simultaneously active units, which are those k units that are most highly activated. In their model, this k value is set low when items are being selected for representation in short term memory, but is then set high for maintenance processes. Representational overlap imposes sharp capacity limitations on how many things can be maintained at any given time, but conversely is also necessary for flexible generalization. This system may therefore be tuned to maximize the competing values in this computational-tradeoff.

Although challenges remain for nonlinear oscillatory models of short term memory capacity limits (such as how oscillations would select the same representations during each "theta cycle," whether that theta cycle is a cause, direct product, or byproduct of maintenance process), they provide a tantalizing view of how a skill that is essential to intelligence may be implemented neurally.

Related Posts:
Entangled Oscillations
Gamma Synchrony
Anticipation and Synchronization