Limits...
Using data-driven model-brain mappings to constrain formal models of cognition.

Borst JP, Nijboer M, Taatgen NA, van Rijn H, Anderson JR - PLoS ONE (2015)

Bottom Line: Although such mappings can be based on the experience of the modeler or on a reading of the literature, a formal method is preferred to prevent researcher-based biases.We then validated this mapping by applying it to two new datasets with associated models.The new mapping was at least as powerful as an existing mapping that was based on the literature, and indicated where the models were supported by the data and where they have to be improved.

View Article: PubMed Central - PubMed

Affiliation: Carnegie Mellon University, Dept. of Psychology, Pittsburgh, United States of America; University of Groningen, Dept. of Artificial Intelligence, Groningen, the Netherlands.

ABSTRACT
In this paper we propose a method to create data-driven mappings from components of cognitive models to brain regions. Cognitive models are notoriously hard to evaluate, especially based on behavioral measures alone. Neuroimaging data can provide additional constraints, but this requires a mapping from model components to brain regions. Although such mappings can be based on the experience of the modeler or on a reading of the literature, a formal method is preferred to prevent researcher-based biases. In this paper we used model-based fMRI analysis to create a data-driven model-brain mapping for five modules of the ACT-R cognitive architecture. We then validated this mapping by applying it to two new datasets with associated models. The new mapping was at least as powerful as an existing mapping that was based on the literature, and indicated where the models were supported by the data and where they have to be improved. We conclude that data-driven model-brain mappings can provide strong constraints on cognitive models, and that model-based fMRI is a suitable way to create such mappings.

Show MeSH
Imaging results of the algebra dataset.The top row shows model predictions, the middle row the BOLD responses in the new, data-driven ROIs, and the bottom row the BOLD responses in the original ROIs.
© Copyright Policy
Related In: Results  -  Collection

License
getmorefigures.php?uid=PMC4352055&req=5

pone.0119673.g006: Imaging results of the algebra dataset.The top row shows model predictions, the middle row the BOLD responses in the new, data-driven ROIs, and the bottom row the BOLD responses in the original ROIs.

Mentions: The top row of Fig. 6 shows the model’s BOLD predictions for four modules: the problem state module (ACT-R’s working memory), declarative memory, the right-manual module, and the visual module. The next two rows show the results in the new, data-driven ROIs and in the original ROIs, respectively. All results are from the ROIs in the left hemisphere, as these typically show the strongest effects. Table 2 reports three fit measures of the model predictions to the data: the R2, the root-mean-square deviation, and Tucker’s Congruence Coefficient (TCC; [61]). Although we assume the first two measures to be familiar, TCC might not be. TCC measures the proportionality of the elements in two vectors, that is, it is another way of measuring the similarity of two vectors. The values of TCC range between −1 and 1, with −1 indicating a complete opposite (with a correlation of −1), and 1 indicating identical vectors. In practice, values between .85 and .94 correspond to a fair similarity, and values over .95 indicate that the two vectors are almost identical [62]. Unlike the R2 measure of correspondence, TCC does take into account the slope of the vectors (positive vs. negative), the sign of the vectors, and it can handle horizontal lines. See S1 Fig. for several demonstrations. In addition to the TCC of the aggregate data, we also report the average TCC per participant, and its standard deviation and range.


Using data-driven model-brain mappings to constrain formal models of cognition.

Borst JP, Nijboer M, Taatgen NA, van Rijn H, Anderson JR - PLoS ONE (2015)

Imaging results of the algebra dataset.The top row shows model predictions, the middle row the BOLD responses in the new, data-driven ROIs, and the bottom row the BOLD responses in the original ROIs.
© Copyright Policy
Related In: Results  -  Collection

License
Show All Figures
getmorefigures.php?uid=PMC4352055&req=5

pone.0119673.g006: Imaging results of the algebra dataset.The top row shows model predictions, the middle row the BOLD responses in the new, data-driven ROIs, and the bottom row the BOLD responses in the original ROIs.
Mentions: The top row of Fig. 6 shows the model’s BOLD predictions for four modules: the problem state module (ACT-R’s working memory), declarative memory, the right-manual module, and the visual module. The next two rows show the results in the new, data-driven ROIs and in the original ROIs, respectively. All results are from the ROIs in the left hemisphere, as these typically show the strongest effects. Table 2 reports three fit measures of the model predictions to the data: the R2, the root-mean-square deviation, and Tucker’s Congruence Coefficient (TCC; [61]). Although we assume the first two measures to be familiar, TCC might not be. TCC measures the proportionality of the elements in two vectors, that is, it is another way of measuring the similarity of two vectors. The values of TCC range between −1 and 1, with −1 indicating a complete opposite (with a correlation of −1), and 1 indicating identical vectors. In practice, values between .85 and .94 correspond to a fair similarity, and values over .95 indicate that the two vectors are almost identical [62]. Unlike the R2 measure of correspondence, TCC does take into account the slope of the vectors (positive vs. negative), the sign of the vectors, and it can handle horizontal lines. See S1 Fig. for several demonstrations. In addition to the TCC of the aggregate data, we also report the average TCC per participant, and its standard deviation and range.

Bottom Line: Although such mappings can be based on the experience of the modeler or on a reading of the literature, a formal method is preferred to prevent researcher-based biases.We then validated this mapping by applying it to two new datasets with associated models.The new mapping was at least as powerful as an existing mapping that was based on the literature, and indicated where the models were supported by the data and where they have to be improved.

View Article: PubMed Central - PubMed

Affiliation: Carnegie Mellon University, Dept. of Psychology, Pittsburgh, United States of America; University of Groningen, Dept. of Artificial Intelligence, Groningen, the Netherlands.

ABSTRACT
In this paper we propose a method to create data-driven mappings from components of cognitive models to brain regions. Cognitive models are notoriously hard to evaluate, especially based on behavioral measures alone. Neuroimaging data can provide additional constraints, but this requires a mapping from model components to brain regions. Although such mappings can be based on the experience of the modeler or on a reading of the literature, a formal method is preferred to prevent researcher-based biases. In this paper we used model-based fMRI analysis to create a data-driven model-brain mapping for five modules of the ACT-R cognitive architecture. We then validated this mapping by applying it to two new datasets with associated models. The new mapping was at least as powerful as an existing mapping that was based on the literature, and indicated where the models were supported by the data and where they have to be improved. We conclude that data-driven model-brain mappings can provide strong constraints on cognitive models, and that model-based fMRI is a suitable way to create such mappings.

Show MeSH