Chris Baldassano - Revision history

imported>Psych204B: /* Data Analysis */

2010-03-19T09:28:31Z

Data Analysis

← Older revision		Revision as of 09:28, 19 March 2010
Line 28:		Line 28:
	[I obtained data that had already been preprocessed using AFNI to provide motion correction and to subtract out the temporal mean of each voxel]		[I obtained data that had already been preprocessed using AFNI to provide motion correction and to subtract out the temporal mean of each voxel]

	First, a spherical searchlight of 81 voxels was centered at every voxel in the brain. Searchlights that did not contain 81 valid voxels (due to being near the edge of the brain, for example) were discarded. An SVM was then trained for each searchlight using the MATLAB implementation of LIBSVM [7]. Each of the 12 runs was held out for cross-validation, one at a time; the accuracy of the SVM was calculated as the average accuracy over these 12 cases. This step was very computationally intensive, requiring approximately 48 hours on a 2.7GHz machine. Searchlight ROIs were then ranked by accuracy.		First, a spherical searchlight of 81 voxels was centered at every voxel in the brain. Searchlights that did not contain 81 valid voxels (due to being near the edge of the brain, for example) were discarded. An SVM was then trained for each searchlight using the MATLAB implementation of LIBSVM [7]. Each of the 12 runs was held out for cross-validation, one at a time; the accuracy of the SVM was calculated as the average accuracy over these 12 cases. This step was very computationally intensive, requiring approximately 48 hours on a 2.7GHz machine (per subject). Searchlight ROIs were then ranked by accuracy.

	A simple form of nonmaximum suppression was used. Any searchlight that overlapped a searchlight with higher accuracy was discarded. The top five ROIs that remained were then selected.		A simple form of nonmaximum suppression was used. Any searchlight that overlapped a searchlight with higher accuracy was discarded. The top five ROIs that remained were then selected.

imported>Psych204B: /* Data Analysis */

2010-03-19T09:28:06Z

Data Analysis

← Older revision		Revision as of 09:28, 19 March 2010
Line 28:		Line 28:
	[I obtained data that had already been preprocessed using AFNI to provide motion correction and to subtract out the temporal mean of each voxel]		[I obtained data that had already been preprocessed using AFNI to provide motion correction and to subtract out the temporal mean of each voxel]

	First, a spherical searchlight of 81 voxels was centered at every voxel in the brain. Searchlights that did not contain 81 valid voxels (due to being near the edge of the brain, for example) were discarded. An SVM was then trained for each searchlight using the MATLAB implementation of LIBSVM [7]. Each of the 12 runs was held out for cross-validation, one at a time; the accuracy of ~~each~~ SVM was calculated as the average accuracy over these 12 cases. This step was very computationally intensive, requiring approximately 48 hours on a 2.7GHz machine. Searchlight ROIs were then ranked by accuracy.		First, a spherical searchlight of 81 voxels was centered at every voxel in the brain. Searchlights that did not contain 81 valid voxels (due to being near the edge of the brain, for example) were discarded. An SVM was then trained for each searchlight using the MATLAB implementation of LIBSVM [7]. Each of the 12 runs was held out for cross-validation, one at a time; the accuracy of the SVM was calculated as the average accuracy over these 12 cases. This step was very computationally intensive, requiring approximately 48 hours on a 2.7GHz machine. Searchlight ROIs were then ranked by accuracy.

	A simple form of nonmaximum suppression was used. Any searchlight that overlapped a searchlight with higher accuracy was discarded. The top five ROIs that remained were then selected.		A simple form of nonmaximum suppression was used. Any searchlight that overlapped a searchlight with higher accuracy was discarded. The top five ROIs that remained were then selected.

imported>Psych204B: /* Data Acquisition */

2010-03-19T09:27:17Z

Data Acquisition

← Older revision		Revision as of 09:27, 19 March 2010
Line 19:		Line 19:
	= Methods =		= Methods =
	=== Data Acquisition ===		=== Data Acquisition ===
	The data used is from the experiment described in [6]. Subjects passively viewed color images of six types of natural scenes: beaches, buildings, forests, highways, industry, and mountains). The stimuli were presented in a block design, with each block composed of 10 images from the same category displayed for 1.6 seconds each (8 brain acquisitions were made during each block). The subject viewed 12 runs of images, with each run composed of six blocks separated by 12 second fixation periods. (The original study also collected 12 runs with inverted images, but this data was not used for the current experiment).		The data used is from the experiment described in [6]. Subjects passively viewed color images of six types of natural scenes: beaches, buildings, forests, highways, industry, and mountains. The stimuli were presented in a block design, with each block composed of 10 images from the same category displayed for 1.6 seconds each (8 brain acquisitions were made during each block). The subject viewed 12 runs of images, with each run composed of six blocks separated by 12 second fixation periods. (The original study also collected 12 runs with inverted images, but this data was not used for the current experiment).

	<gallery caption="Example Scene Images" widths="500px" heights="400px" perrow="1">		<gallery caption="Example Scene Images" widths="500px" heights="400px" perrow="1">

imported>Psych204B: /* Scene Processing */

2010-03-19T09:25:50Z

Scene Processing

← Older revision		Revision as of 09:25, 19 March 2010
Line 4:		Line 4:
	= Background =		= Background =
	== Scene Processing ==		== Scene Processing ==
	The dataset being used for this project is designed to investigate natural scene perception (see Data Acquisition below). This research area has begun to attract significant attention, since humans display an incredible proficiency at classifying natural scenes: recognition can occur with presentation times as short as 100 ms [2] and can access many details about scenes in only a glance [3]. This efficiency places constraints on the neural basis of scene perception, and suggests that modeling this process may be tractable.		The dataset being used for this project is designed to investigate natural scene perception (see Data Acquisition below). This research area has begun to attract significant attention, since humans display an incredible proficiency at classifying natural scenes: recognition can occur with presentation times as short as 100 ms [2] and subjects can access many details about scenes in only a glance [3]. This efficiency places constraints on the neural basis of scene perception, and suggests that modeling this process may be tractable.

	== Support Vector Machines ==		== Support Vector Machines ==

imported>Psych204B: /* Complementary Connectivity */

2010-03-19T09:24:54Z

Complementary Connectivity

← Older revision		Revision as of 09:24, 19 March 2010
Line 11:		Line 11:
	== Complementary Connectivity ==		== Complementary Connectivity ==
	The connectivity analysis for this project was conducted based on the graphical model in [5], as shown in Figure 1. The probability estimates output from each SVM are the X variables. Rather than having each ROI "vote" independently to predict the scene type, the predictions from multiple ROIs are allowed to interact in the middle layer. If a connection between two of the Y variables improves classification performance, this indicates that the connected ROIs have complementary information about the scene. Connectivity is learned by comparing classification performance with all the Ys unconnected and with one pair of Ys connected. All of the pairwise connections that improve performance are then added to the final connectivity graph. Although this procedure is not guaranteed to find the minimal connectivity graph, it decreases the running time from <math>O(2^{n^2})</math> to <math>O(n^2)</math>.		The connectivity analysis for this project was conducted based on the graphical model in [5], as shown in Figure 1. The probability estimates output from each SVM are the X variables. Rather than having each ROI "vote" independently to predict the scene type, the predictions from multiple ROIs are allowed to interact in the middle layer. If a connection between two of the Y variables improves classification performance, this indicates that the connected ROIs have complementary information about the scene. Connectivity is learned by comparing classification performance with all the Ys unconnected and with one pair of Ys connected. All of the pairwise connections that improve performance are then added to the final connectivity graph. Although this procedure is not guaranteed to find the minimal connectivity graph, it decreases the running time from <math>O(2^{n^2})</math> to <math>O(n^2)</math>.

			Note that in order to evaluate the performance of a given graph structure, it is necessary to learn the edge weights of the graph. Unfortunately, there is no closed-form expression for this maximization, so iterative methods are required. Even more unfortunately, the objective function is not concave, so the edge weights we learn in an iterative algorithm may depend on the initial values with which we seed the algorithm. The approach taken in this case is to simply run gradient ascent with different starting values, and observe which graph structures frequently give good performance.
	<gallery caption="" widths="700px" heights="400px" perrow="1">		<gallery caption="" widths="700px" heights="400px" perrow="1">
	File:Connectivitymodel.jpg \|Figure 1		File:Connectivitymodel.jpg \|Figure 1

imported>Psych204B: /* Conclusions */

2010-03-19T09:21:00Z

Conclusions

← Older revision		Revision as of 09:21, 19 March 2010
Line 317:		Line 317:
	Using a searchlight based on SVM accuracy successfully located brain areas known to be involved in scene processing. Some of the ROIs selected, however, could not be easily mapped to visual areas. These areas can be studied further to identify the role that they play in scene processing.		Using a searchlight based on SVM accuracy successfully located brain areas known to be involved in scene processing. Some of the ROIs selected, however, could not be easily mapped to visual areas. These areas can be studied further to identify the role that they play in scene processing.

	Using an SVM searchlight is a general MVPA approach that could be applied in many experiments to identify brain areas that are informative about the stimulus. Training using SVMs is simple, since robust libraries exist that can apply complex optimization algorithms to the data. One disadvantage of this method is its computational complexity. For this study, we performed 12-fold cross-validation at every voxel, requiring us to train a total of 12*(~20,000) = 240,000 SVM classifiers for every subject. Although training a single SVM is relatively efficient (~1 second on my 2.7GHz desktop), training this many classifiers requires a large amount of computation. For a larger study, a computing cluster would be required. Note that each classifier is trained independently, so this process can very easily by parallelized over multiple machines.		Using an SVM searchlight is a general MVPA approach that could be applied in many experiments to identify brain areas that are informative about the stimulus. Training using SVMs is simple, since robust libraries exist that can apply complex optimization algorithms to the data. One disadvantage of this method is its computational complexity. For this study, we performed 12-fold cross-validation at every voxel, requiring us to train a total of 12*(~20,000) = 240,000 SVM classifiers for every subject. Although training a single SVM is relatively efficient (~1 second on my 2.7GHz desktop), training this many classifiers requires a large amount of computation. For a larger study, a computing cluster would be required. Note that each classifier is trained independently, so this process can very easily be parallelized over multiple machines.

	The connectivity results are difficult to interpret, since the learned connections depend heavily on the initial settings of the parameters. Although I achieved reasonably consistent results, a better approach would involve a much more extensive search of the initial parameter space (e.g. using different initial weight values for different edges). Again, this would take prohibitively long on a single computer, but could easily by parallelized over many computers.		The connectivity results are difficult to interpret, since the learned connections depend heavily on the initial settings of the parameters. Although I achieved reasonably consistent results, a better approach would involve a much more extensive search of the initial parameter space (e.g. using different initial weight values for different edges). Again, this would take prohibitively long on a single computer, but could easily by parallelized over many computers.

imported>Psych204B: /* Conclusions */

2010-03-19T09:20:12Z

Conclusions

← Older revision		Revision as of 09:20, 19 March 2010
Line 315:		Line 315:
	= Conclusions =		= Conclusions =

	Using a searchlight based on SVM accuracy successfully located brain areas known to be involved in scene processing. Some of the ROIs selected, however, could not be easily mapped to visual areas. These areas can be studied further to identify the role that they play in scene processing~~. Using an SVM searchlight is a general MVPA approach that could be applied in many experiments to identify brain areas that are informative about the stimulus~~.		Using a searchlight based on SVM accuracy successfully located brain areas known to be involved in scene processing. Some of the ROIs selected, however, could not be easily mapped to visual areas. These areas can be studied further to identify the role that they play in scene processing.

	The connectivity results are difficult to interpret, since the learned connections depend heavily on the initial settings of the parameters. Although I achieved reasonably consistent results, a better approach would involve a much more extensive search of the initial parameter space (e.g. using different initial weight values for different edges).		Using an SVM searchlight is a general MVPA approach that could be applied in many experiments to identify brain areas that are informative about the stimulus. Training using SVMs is simple, since robust libraries exist that can apply complex optimization algorithms to the data. One disadvantage of this method is its computational complexity. For this study, we performed 12-fold cross-validation at every voxel, requiring us to train a total of 12*(~20,000) = 240,000 SVM classifiers for every subject. Although training a single SVM is relatively efficient (~1 second on my 2.7GHz desktop), training this many classifiers requires a large amount of computation. For a larger study, a computing cluster would be required. Note that each classifier is trained independently, so this process can very easily by parallelized over multiple machines.

			The connectivity results are difficult to interpret, since the learned connections depend heavily on the initial settings of the parameters. Although I achieved reasonably consistent results, a better approach would involve a much more extensive search of the initial parameter space (e.g. using different initial weight values for different edges). Again, this would take prohibitively long on a single computer, but could easily by parallelized over many computers.

	= References =		= References =

imported>Psych204B: /* Connectivity Results */

2010-03-16T01:24:27Z

Connectivity Results

← Older revision		Revision as of 01:24, 16 March 2010
Line 311:		Line 311:

	One interesting result is that ROI #1 for subject W5 is by far the most strongly connected to other ROIs. Since we identified this ROI as part of RSC, this matches findings in [8] and [5] that RSC plays an important role in scene processing.		One interesting result is that ROI #1 for subject W5 is by far the most strongly connected to other ROIs. Since we identified this ROI as part of RSC, this matches findings in [8] and [5] that RSC plays an important role in scene processing.
			We also see that PPA in subject B7 (ROI #2) has the strongest complementary connections to other visual areas.

	= Conclusions =		= Conclusions =

imported>Psych204B: /* Searchlight Results */

2010-03-16T00:57:53Z

Searchlight Results

← Older revision		Revision as of 00:57, 16 March 2010
Line 73:		Line 73:
	These values are all well above the chance level of 1/6 = 16.7%. We can observe that the accuracy has a relatively strong dependence on the specific subject; this could be due to different noise parameters during different scans, or higher-level attentional differences.		These values are all well above the chance level of 1/6 = 16.7%. We can observe that the accuracy has a relatively strong dependence on the specific subject; this could be due to different noise parameters during different scans, or higher-level attentional differences.

	The physical locations of the selected ROIs are shown in Figures 3-5. The ROIs for each subject are presented in increasing Z slice location, with numbers indicating their relative SVM accuracy (1 = most accurate). We can identify a number of visual-processing areas among these ROIs. For example, we have identified PPA (C4#1 and W5#5), RSC (W5#1), V1(B7#1) and Cuneus (B7#5 and C4#3).		The physical locations of the selected ROIs are shown in Figures 3-5. The ROIs for each subject are presented in increasing Z slice location, with numbers indicating their relative SVM accuracy (1 = most accurate). We can identify a number of visual-processing areas among these ROIs. For example, we have identified PPA (B7#2, C4#1 and W5#5), RSC (W5#1), V1(B7#1) and Cuneus (B7#5 and C4#3).

	<gallery caption="Searchlight Results" widths="850px" heights="240px" perrow="1">		<gallery caption="Searchlight Results" widths="850px" heights="240px" perrow="1">

imported>Psych204B: /* Conclusions */

2010-03-16T00:52:18Z

Conclusions

← Older revision		Revision as of 00:52, 16 March 2010
Line 314:		Line 314:
	= Conclusions =		= Conclusions =

	~~Here~~ is ~~where you say what your~~ results ~~mean~~.		Using a searchlight based on SVM accuracy successfully located brain areas known to be involved in scene processing. Some of the ROIs selected, however, could not be easily mapped to visual areas. These areas can be studied further to identify the role that they play in scene processing. Using an SVM searchlight is a general MVPA approach that could be applied in many experiments to identify brain areas that are informative about the stimulus.

			The connectivity results are difficult to interpret, since the learned connections depend heavily on the initial settings of the parameters. Although I achieved reasonably consistent results, a better approach would involve a much more extensive search of the initial parameter space (e.g. using different initial weight values for different edges).

	= References =		= References =