Limits...
Inferring functional brain states using temporal evolution of regularized classifiers.

Zhdanov A, Hendler T, Ungerleider L, Intrator N - Comput Intell Neurosci (2007)

Bottom Line: We approach the problem from a machine learning perspective, by constructing a classifier from a set of labeled signal examples.We propose a framework that focuses on temporal evolution of regularized classifiers, with cross-validation for optimal regularization parameter at each time frame.We demonstrate the inference obtained by this method on MEG data recorded from 10 subjects in a simple visual classification experiment, and provide comparison to the classical nonregularized approach.

View Article: PubMed Central - PubMed

Affiliation: Functional Brain Imaging Unit, Tel Aviv Sourasky Medical Center, 6 Weizmann Street, Tel Aviv 64239, Israel. zhdanova@post.tau.ac.il

ABSTRACT
We present a framework for inferring functional brain state from electrophysiological (MEG or EEG) brain signals. Our approach is adapted to the needs of functional brain imaging rather than EEG-based brain-computer interface (BCI). This choice leads to a different set of requirements, in particular to the demand for more robust inference methods and more sophisticated model validation techniques. We approach the problem from a machine learning perspective, by constructing a classifier from a set of labeled signal examples. We propose a framework that focuses on temporal evolution of regularized classifiers, with cross-validation for optimal regularization parameter at each time frame. We demonstrate the inference obtained by this method on MEG data recorded from 10 subjects in a simple visual classification experiment, and provide comparison to the classical nonregularized approach.

No MeSH data available.


Error rate asa function of regularization parameter for subject ZK. Solid blue line denotesthe average error rate over 100-fold cross-validation, dotted lines mark1-std-wide margin; the vertical line marks the minimum of the smoothed errorrate (red line). Three plots below show the distribution of sensor weightscorresponding to different values of the regularization parameter.
© Copyright Policy - open-access
Related In: Results  -  Collection


getmorefigures.php?uid=PMC2266829&req=5

fig5: Error rate asa function of regularization parameter for subject ZK. Solid blue line denotesthe average error rate over 100-fold cross-validation, dotted lines mark1-std-wide margin; the vertical line marks the minimum of the smoothed errorrate (red line). Three plots below show the distribution of sensor weightscorresponding to different values of the regularization parameter.

Mentions: We examined the weight maps obtained for the combinationof λ and timeslicethat yield the lowest estimated prediction error. The maps display a prominentstructure consisting of several small clusters of interleaved positive andnegative weights (see Figure 4). As expected from animal single unit and fMRIhuman studies [27],this structure is fairly localized to occipitotemporal regions that mightcorrespond to a neural source in the fusiform gyrus. The structure seems to bemore clearly exhibited in the predictable subjects. We also investigated therelation between the value of λ and thestructure of corresponding weight maps. As one could have expected, increasingthe regularization parameter causes the resulting optimal weight maps to becomesmoother (see Figure 5).


Inferring functional brain states using temporal evolution of regularized classifiers.

Zhdanov A, Hendler T, Ungerleider L, Intrator N - Comput Intell Neurosci (2007)

Error rate asa function of regularization parameter for subject ZK. Solid blue line denotesthe average error rate over 100-fold cross-validation, dotted lines mark1-std-wide margin; the vertical line marks the minimum of the smoothed errorrate (red line). Three plots below show the distribution of sensor weightscorresponding to different values of the regularization parameter.
© Copyright Policy - open-access
Related In: Results  -  Collection

Show All Figures
getmorefigures.php?uid=PMC2266829&req=5

fig5: Error rate asa function of regularization parameter for subject ZK. Solid blue line denotesthe average error rate over 100-fold cross-validation, dotted lines mark1-std-wide margin; the vertical line marks the minimum of the smoothed errorrate (red line). Three plots below show the distribution of sensor weightscorresponding to different values of the regularization parameter.
Mentions: We examined the weight maps obtained for the combinationof λ and timeslicethat yield the lowest estimated prediction error. The maps display a prominentstructure consisting of several small clusters of interleaved positive andnegative weights (see Figure 4). As expected from animal single unit and fMRIhuman studies [27],this structure is fairly localized to occipitotemporal regions that mightcorrespond to a neural source in the fusiform gyrus. The structure seems to bemore clearly exhibited in the predictable subjects. We also investigated therelation between the value of λ and thestructure of corresponding weight maps. As one could have expected, increasingthe regularization parameter causes the resulting optimal weight maps to becomesmoother (see Figure 5).

Bottom Line: We approach the problem from a machine learning perspective, by constructing a classifier from a set of labeled signal examples.We propose a framework that focuses on temporal evolution of regularized classifiers, with cross-validation for optimal regularization parameter at each time frame.We demonstrate the inference obtained by this method on MEG data recorded from 10 subjects in a simple visual classification experiment, and provide comparison to the classical nonregularized approach.

View Article: PubMed Central - PubMed

Affiliation: Functional Brain Imaging Unit, Tel Aviv Sourasky Medical Center, 6 Weizmann Street, Tel Aviv 64239, Israel. zhdanova@post.tau.ac.il

ABSTRACT
We present a framework for inferring functional brain state from electrophysiological (MEG or EEG) brain signals. Our approach is adapted to the needs of functional brain imaging rather than EEG-based brain-computer interface (BCI). This choice leads to a different set of requirements, in particular to the demand for more robust inference methods and more sophisticated model validation techniques. We approach the problem from a machine learning perspective, by constructing a classifier from a set of labeled signal examples. We propose a framework that focuses on temporal evolution of regularized classifiers, with cross-validation for optimal regularization parameter at each time frame. We demonstrate the inference obtained by this method on MEG data recorded from 10 subjects in a simple visual classification experiment, and provide comparison to the classical nonregularized approach.

No MeSH data available.