Are imaging and lesioning convergent methods for assessing functional specialisation? Investigations using an artificial neural network

This article presents an investigation of the relationship between lesioning and neuroimaging methods of assessing functional specialisation, using synthetic brain imaging (SBI) and lesioning of a connectionist network of past-tense formation. The model comprised two processing ‘routes’: one was a direct route between layers of input and output units, while the other, indirect, route featured an intermediate layer of processing units. Emergent specialisation within the network was assessed (1) by lesioning either the direct or indirect route and measuring past-tense performance for regular and irregular verbs, and (2) by measuring functional activation in each route when processing each verb type (SBI). SBI and lesioning approaches failed to converge when network activation was summed over each route in our SBI approach. Examination of individual network solutions suggested that the verb types might be using the indirect route differently in terms of the pattern of activation across the route, rather than in terms of gross activation. A subsequent SBI analysis compared patterns of activation in the indirect route and confirmed that these patterns were more similar between regular-type verbs than between regular and irregular verbs. As the spatial and temporal resolution of neuroimaging techniques improves, the results of this investigation suggest that the key to finding functional specialisation will be to distinguish local coding differences across behaviours that are the results of developmental processes. Other analyses suggest that lesioning data may be limited because, with increasing damage, they reveal the resting activations of a computational system rather than a computational specialisation per se.

► We use computational modelling to contrast lesioning and functional neuroimaging.
► The model shows how fMRI of gross brain regions may be blind to functional specialisation.
► Differing patterns of neural coding within a region may instead reveal specialisation.
► Lesioning violates the transparency assumption.
► The model’s functional specialisation was not pre-wired, but acquired developmentally.