Functional magnetic resonance imaging examination of two modular architectures for switching multiple internal models - PubMed (original) (raw)

Functional magnetic resonance imaging examination of two modular architectures for switching multiple internal models

Hiroshi Imamizu et al. J Neurosci. 2004.

Abstract

An internal model is a neural mechanism that can mimic the input-output properties of a controlled object such as a tool. Recent research interests have moved on to how multiple internal models are learned and switched under a given context of behavior. Two representative computational models for task switching propose distinct neural mechanisms, thus predicting different brain activity patterns in the switching of internal models. In one model, called the mixture-of-experts architecture, switching is commanded by a single executive called a "gating network," which is different from the internal models. In the other model, called the MOSAIC (MOdular Selection And Identification for Control), the internal models themselves play crucial roles in switching. Consequently, the mixture-of-experts model predicts that neural activities related to switching and internal models can be temporally and spatially segregated, whereas the MOSAIC model predicts that they are closely intermingled. Here, we directly examined the two predictions by analyzing functional magnetic resonance imaging activities during the switching of one common tool (an ordinary computer mouse) and two novel tools: a rotated mouse, the cursor of which appears in a rotated position, and a velocity mouse, the cursor velocity of which is proportional to the mouse position. The switching and internal model activities temporally and spatially overlapped each other in the cerebellum and in the parietal cortex, whereas the overlap was very small in the frontal cortex. These results suggest that switching mechanisms in the frontal cortex can be explained by the mixture-of-experts architecture, whereas those in the cerebellum and the parietal cortex are explained by the MOSAIC model.

PubMed Disclaimer

Figures

Figure 1.

Figure 1.

Two representative computational methods for switching internal models. A, Mixture-of-experts architecture: a gating module estimates the current context and decides the contribution of each expert (internal model) independently of the activity of the experts. Switching functions (hatched parts) are concentrated in the gating module. B, MOSAIC architecture: a forward model and an inverse model that are tightly coupled, and the RE decides the contribution of each pair of internal models according to the goodness of prediction made by the forward models. Internal models contribute to the switching functions.

Figure 5.

Figure 5.

Time courses of activation (relative BOLD signal; see Materials and Methods) in regions related to the switching of the mouse type. The activated regions are the same as those in Figure 4 A (random effect model; t(9) > 4.3; p < 0.001 uncorrected; cluster size, > 50 voxels) but shown in transverse sections. The time courses were averaged across repetitions, subjects, and voxels in each region. The abscissa represents time from the switching of mouse type in seconds. The color of lines and markers indicates the mouse type being used at each time point (orange, rotation; blue, velocity; black, normal). As illustrated in the bottom left corner, the solid lines with open circles indicate time courses when the mouse type changed from the normal mouse to the rotated mouse (the left figure in each section) or velocity mouse (the right figure). The broken lines with crosses indicate those when the mouse type changed from the rotated mouse (the left figure) or velocity mouse (the right figure) to the normal mouse. The time course marked with an arrow (E) will be used as an example in Figure 6.

Figure 6.

Figure 6.

An example for quantitative analysis of the activation time courses to separate and compare the sustained component and transient component. A solid line with open circles indicates one of the time courses observed in the right cerebellum (marked with an arrow in Fig. 5_E_). The time course was fitted using a general linear model by a weighted sum of three components: a step function modeling the sustained activity (magenta curve); a pulse function modeling the transient activity (green curve); and a constant component. The height of the curve peaks from the baseline (p and q) indicates the estimated weights of the sustained component and the transient component. A broken curve represents the summation of the three estimated components (a fitted time course).

Figure 2.

Figure 2.

Tracking errors (across-subjects; mean ± SD) when subjects manipulate the rotated mouse (solid lines with filled circles), velocity mouse (broken lines with open circles), and normal mouse (triangles) in training sessions.

Figure 3.

Figure 3.

Time courses of errors during scanning sessions for four types of transition between blocks. The time courses were aligned on the transition. The markers (circles and triangles) and bars indicate across-subjects mean ± SD at every second. The asterisk indicates a significant difference in the errors between when the mouse type was changed and when the type was not changed (p < 0.05; multiple comparisons using Student's t test).

Figure 4.

Figure 4.

Regions related to the switching of the mouse type (A), cognitive cues (B), and cursor reset (C) (random effect model; t(9) > 4.3; p < 0.001 uncorrected; cluster size, >50 voxels).

Figure 7.

Figure 7.

Activations related to the behavioral switch (red), rotated mouse (orange), and velocity mouse (blue) in the parietal regions (A), the PM regions (B), and the cerebellum (C) (random effect model; t(9) > 4.3; p < 0.001 uncorrected; cluster size, >50 voxels). The illustrations above the activation maps indicate the VOI and the percentage of overlap between the switching-related volume and the volume related to the rotated or velocity mouse (mosaic of red and orange, or red and blue) to the switching-related volume (red) within each VOI. The illustrations in A and B show the horizontal sections, whereas the illustration in C shows the coronal section. All activation maps show transverse sections.

Figure 8.

Figure 8.

Two possible schemata regarding the switch mechanisms of internal models implicated by the current results. A assumes that the switching is conducted only by the MOSAIC architecture. B assumes that it is conducted by the mixture-of-experts and the MOSAIC depending on task requirements (see Discussion). The blue parts are related to the internal models, whereas the red parts are related to the switching functions. RP, Responsibility predictor.

Similar articles

Cited by

References

    1. Bischoff-Grethe A, Ivry RB, Grafton ST (2002) Cerebellar involvement in response reassignment rather than attention. J Neurosci 22: 546–553. - PMC - PubMed
    1. Blakemore SJ, Wolpert DM, Frith CD (1998) Central cancellation of self-produced tickle sensation. Nat Neurosci 1: 635–640. - PubMed
    1. Clower DM, West RA, Lynch JC, Strick PL (2001) The inferior parietal lobule is the target of output from the superior colliculus, hippocampus, and cerebellum. J Neurosci 21: 6283–6291. - PMC - PubMed
    1. D'Esposito M, Detre JA, Alsop DC, Shin RK, Atlas S, Grossman M (1995) The neural basis of the central executive system of working memory. Nature 378: 279–281. - PubMed
    1. Desmond JE, Fiez JA (1998) Neuroimaging studies of the cerebellum: language, learning and memory. Trends Cogn Sci 2: 355–362. - PubMed

Publication types

MeSH terms

LinkOut - more resources