Introduction

 fMRI intro
Functional Magnetic Resonance Imaging (fMRI) is a technique for scanning the brain in three dimensions and determining the elements present at each point (a three-dimensional space called a "voxel"). By measuring the blood oxygen level deficiency (BOLD) signal, inferences can be made about neuronal activity at different parts of the brain. The “functional” descriptor of fMRI indicates that these images are taken with the purpose of associating neuronal activity with different functions of the brain, as assessed while subjects perform certain tasks. Its non-invasive quality has made fMRI the dominant method for studies of the function of the human brain, being frequently used to understand the neurobiological underpinnings of various human tasks and processes.
 
studyforrest Dataset and Annotations
The time and resources required to obtain fMRI data make large, publicly available datasets invaluable to the furtherance of neuropsychological research. One such example is the studyforrest dataset, originally collected on 20 individuals as they listened to the audio descriptive version of the movie Forrest Gump. Later expansions to the dataset included 3T fMRI as well as eye-tracking data for 15 of the original subjects as they watched the audio-visual version of the movie (CITE). The University of Magdeburg team responsible for the data published it with the intent of enabling the study of auditory attention; cognition; and language, music, and social perception.
Annotations of Forrest Gump have also been made available to enable more detailed analyses of the studyforrest dataset, such as that published by Labs et al. (CITE), which details episodes of portrayed emotion across both the audio description and audiovisual versions of the movie. Data on emotions together with arousal, valence, and direction classifications, was obtained from multiple raters and is available both in raw form and in a consensus, keeping those time windows with a minimum of at least 50% inter-observer agreement.
Applications of fMRI Machine Learning
Machine learning analyses are often performed using fMRI data to “decode” brain responses to certain stimuli. \citet{Ahmad_2016} provide a basic example of this, using a support vector machine (SVM) to predict the cognitive state of subjects, yielding 70.2% accuracy. Data processing, brain segmentation, and feature selection all have the potential to greatly increase accuracy, as shown in the following studies.
 
\citet{Casey_2017}, in their paper on decoding multivoxel responses to various musical stimuli, exemplify two different paradigms for machine learning prediction on fMRI data. In the more basic Multivariate Pattern Analysis (MVPA) approach, each voxel is a feature. Because fMRI data typically measures thousands or tens of thousands of voxels, there are too many voxels to treat this way and they need to be reduced in number to avoid the curse of dimensionality and need to be reduced, often by region of interest (ROI) specification. This was accomplished using what in computational anatomy is called an “atlas.” Casey et al., in this case, in addition kept only a subset of the voxels in each ROI before employing a support vector machine (SVM) classifier. In this paper, they also employed the “searchlight” method, which considers not individual voxels, but spheres of voxels of specified radius.
Smyser et al. (CITE) perform machine learning on prematurely-born infants fMRI data using the MVPA method. First they transform the data to account for differing brain sizes, limit it to 100 frames, and divide the voxels into 214 gray-matter “parcels.” They then use machine learning software called Spider to perform SVM analyses on the data which they validated using LOOCV (leave one out cross validation). Prediction accuracy reached up to 84%, and was not seriously affected after performing subset analyses to avoid confounding.
 Mansson et al. (CITE) provide another example of using machine learning on fMRI data to make predictions—in this case on the long-term response of individuals receiving treatments SAD (social anxiety disorder). ROIs were selected that were involved with the fear network: the ACC (anterior cingulate cortex), amygdala, hippocampus, and insula. Using the Pattern Recognition for Neuroimaging Toolbox, delimited ROIs and ran SVM analyses. The ACC turned out to be most informative, being able to predict with 91.7% accuracy the subjects who would respond 1 year after treatment.
Advanced Dimensionality Reduction
Despite the apparent popularity of MVPA, Michel et al. (CITE) present a supervised clustering approach for brain segmentation. In their method, they use a technique called “Ward clustering” to create a hierarchically clustered tree of voxels, which they then prune using a predictive value function—thus maximizing prediction accuracy.  The optimal subtree is then selected from this set of nested parcellations, also for maximum predictive value. This method provides greater accuracy than standard SVMs and is fairly interpretable as well.
Brain's processing of emotions
Previous attempts to study the neurobiology of emotion perception include a study performed by Grezes et al.\cite{Gr_zes_2007}, who performed fMRI analysis of subjects as they observed both images and videos of people either emotionally neutral or portraying fear. After performing statistical analyses on activation of different brain regions, they found that seeing portrayals of fear activated the amygdala, as well as in the temporal pole and lateral orbital cortex. These findings agree with previous studies, as these three regions had been identified before to be important in the perception of fear. They also found that the right premotor cortex and right temporal sulcus could be related to perceptions of dynamic (as opposed to static) portrayals of fear.
 
 

Prop

Data

Given fMRI data of individuals watching the movie Forrest Gump, we will attempt to predict the emotion the individual is viewing at the time.

f

This data was collected by independent observes who did not collaborate when making their annotations. This means data from the different observers must be combined in some manner.
The authors equip the annotation with a python script that creates two sets of data files we may draw from. The first set divides the movie into 1 second intervals, and for each character, gives a percentage of raters that stated a character had a particular emotion in that instance; this is then repeated for 2 second intervals (ie, the length of time in each voxel), and whole scenes. The second set takes creates a threshold for each data point in the first set, and includes data points for each character that has emotional data that cross that threshold.
This leaves us with three sets of data: A raw data folder, which contains a file for every rater-character combination, in which there exists ratings for every second for that rater; character average data, which contains a file for each character with average values of each emotional feature for each timeframe, and threshold data, a single file that contains all points from the the character-average data that had an attribute cross a particular threshold. While we
will only use this final dataset for creating labels, experimenting with threshold values in the character-average set may also be necessary, if the default data set does not produce results.

Methods

Creation of correct label data

We must first filter the fMRI data for voxels within the time frame in each the the data producer’s python script produces label data above threshold, for each individual subject. This can be done with R’s dplyr package.
Our use of labels derived from viewing a single movie must deal with two issues specific to the nature of the dataset. First, the set of emotions in the data set is quite large (n = 22); even given the large amount of fMRI time given to each participant, there will be many labels for which not enough data exist to produce accurate predictions. Second, emotional reactions may vary according to the character portraying the emotion (ie, Forrest’s sadness may not have the effect as Jenny’s sadness). We will attempt to solve this by choosing labels the following ways:

Preprocessing fMRI data

Most preprocessing will be done using the Python package nilearn. fMRI data must first be masked, to remove all sections of each image that are not part of the participant’s brain--especially the face--and to provide basic feature reduction of irrelevant voxels in the image. The fMRI dataset provides a mask for each set of images that may be applied using the nilearn package. Feature reduction may also enhanced by using ANOVA, which is implemented with the sklearn.svm python package. Each image must also first be smoothed and reformed; each individual has slightly different brain proportions, and so the data must be averaged to create an accurate picture of the location of brain activation across several individuals. These
Machine learning tests
Data will be tested using the searchlight method, which applies a data classifier on specific spherical sections to find which sections contain the most relevant data. We will compile data from various learning methods, including linear SVM, logistic regression, and k-nearest neighbor. Each statistical model will be carried out with different parameters; logistic regression with different penalties, and k-nearest neighbor with different values of k.
We will split the development of preprocessing each sets of data (one for fMRI data, and the other for label data); once completed, we will split data gathering between each model (one for linear SVM, one for logistic regression).
The use of R for statistical analysis of fMRI data is tricky. R has a package for reading nifiti images, and can be used to conduct machine learning on  preprocessed image, but has no package for preprocessing fMRI data; it must be done in Python. R also does not have a package for conducting searchlight style fMRI analysis. While searchlight is not necessary for purely predictive analysis, it is useful for determining which areas are involved in producing each emotion.  The preprocessing of data, and preparing data for switching back and forth between languages will take much more time (in terms of effort) than the gathering of data. We also do not generally know the time scale involved for running each algorithm; each run may take an excessive amount of time. Each of these considerations may result in final adjustments in the results we attempt to obtain.

Proposed results

The following images encapsulate our proposed results:
  1. Accuracy of k-nearest neighbor across different values of k;
  2. Accuracy of k-nearest neighbor for each kind of label;
  3. Accuracy of logistic regression for penalty values of 1 and 2;
  4. Accuracy of logistic regression for each kind of label;
  5. Accuracy of linear SVM across different kind of labels.
  6. Searchlight image of high arousal emotions vs low arousal emotions, and for distinct specific emotions.
  7. Accuracy of k-nearest neighbor across different values of k;
  8. Accuracy of k-nearest neighbor for each kind of label;
  9. Accuracy of logistic regression for penalty values of 1 and 2;
  10. Accuracy of logistic regression for each kind of label;
  11. Accuracy of linear SVM across different kind of labels.
  12. Searchlight image of high arousal emotions vs low arousal emotions, and for distinct specific emotions.
  13. Accuracy of k-nearest neighbor across different values of k;
  14. Accuracy of k-nearest neighbor for each kind of label;
  15. Accuracy of logistic regression for penalty values of 1 and 2;
  16. Accuracy of logistic regression for each kind of label;
  17. Accuracy of linear SVM across different kind of labels.
  18. Searchlight image of high arousal emotions vs low arousal emotions, and for distinct specific emotions.
  19. Accuracy of k-nearest neighbor across different values of k;
  20. Accuracy of k-nearest neighbor for each kind of label;
  21. Accuracy of logistic regression for penalty values of 1 and 2;
  22. Accuracy of logistic regression for each kind of label;
  23. Accuracy of linear SVM across different kind of labels.
  24. Searchlight image of high arousal emotions vs low arousal emotions, and for distinct specific emotions.
  25. Accuracy of k-nearest neighbor across different values of k;
  26. Accuracy of k-nearest neighbor for each kind of label;
  27. Accuracy of logistic regression for penalty values of 1 and 2;
  28. Accuracy of logistic regression for each kind of label;
  29. Accuracy of linear SVM across different kind of labels.
  30. Searchlight image of high arousal emotions vs low arousal emotions, and for distinct specific emotions.
  31. Accuracy of k-nearest neighbor across different values of k;
  32. Accuracy of k-nearest neighbor for each kind of label;
  33. Accuracy of logistic regression for penalty values of 1 and 2;
  34. Accuracy of logistic regression for each kind of label;
  35. Accuracy of linear SVM across different kind of labels.
  36. Searchlight image of high arousal emotions vs low arousal emotions, and for distinct specific emotions.
  37. Accuracy of k-nearest neighbor across different values of k;
  38. Accuracy of k-nearest neighbor for each kind of label;
  39. Accuracy of logistic regression for penalty values of 1 and 2;
  40. Accuracy of logistic regression for each kind of label;
  41. Accuracy of linear SVM across different kind of labels.
  42. Searchlight image of high arousal emotions vs low arousal emotions, and for distinct specific emotions.
  43. Accuracy of k-nearest neighbor across different values of k;
  44. Accuracy of k-nearest neighbor for each kind of label;
  45. Accuracy of logistic regression for penalty values of 1 and 2;
  46. Accuracy of logistic regression for each kind of label;
  47. Accuracy of linear SVM across different kind of labels.
  48. Searchlight image of high arousal emotions vs low arousal emotions, and for distinct specific emotions.
  49. Accuracy of k-nearest neighbor across different values of k;
  50. Accuracy of k-nearest neighbor for each kind of label;
  51. Accuracy of logistic regression for penalty values of 1 and 2;
  52. Accuracy of logistic regression for each kind of label;
  53. Accuracy of linear SVM across different kind of labels.
  54. Searchlight image of high arousal emotions vs low arousal emotions, and for distinct specific emotions.
  55. Accuracy of k-nearest neighbor across different values of k;
  56. Accuracy of k-nearest neighbor for each kind of label;
  57. Accuracy of logistic regression for penalty values of 1 and 2;
  58. Accuracy of logistic regression for each kind of label;
  59. Accuracy of linear SVM across different kind of labels.
  60. Searchlight image of high arousal emotions vs low arousal emotions, and for distinct specific emotions.
  61. Accuracy of k-nearest neighbor across different values of k;
  62. Accuracy of k-nearest neighbor for each kind of label;
  63. Accuracy of logistic regression for penalty values of 1 and 2;
  64. Accuracy of logistic regression for each kind of label;
  65. Accuracy of linear SVM across different kind of labels.
  66. Searchlight image of high arousal emotions vs low arousal emotions, and for distinct specific emotions.
  67. Accuracy of k-nearest neighbor across different values of k;
  68. Accuracy of k-nearest neighbor for each kind of label;
  69. Accuracy of logistic regression for penalty values of 1 and 2;
  70. Accuracy of logistic regression for each kind of label;
  71. Accuracy of linear SVM across different kind of labels.
  72. Searchlight image of high arousal emotions vs low arousal emotions, and for distinct specific emotions.
  73. Accuracy of k-nearest neighbor across different values of k;
  74. Accuracy of k-nearest neighbor for each kind of label;
  75. Accuracy of logistic regression for penalty values of 1 and 2;
  76. Accuracy of logistic regression for each kind of label;
  77. Accuracy of linear SVM across different kind of labels.
  78. Searchlight image of high arousal emotions vs low arousal emotions, and for distinct specific emotions.
  79. Accuracy of k-nearest neighbor across different values of k;
  80. Accuracy of k-nearest neighbor for each kind of label;
  81. Accuracy of logistic regression for penalty values of 1 and 2;
  82. Accuracy of logistic regression for each kind of label;
  83. Accuracy of linear SVM across different kind of labels.
  84. Searchlight image of high arousal emotions vs low arousal emotions, and for distinct specific emotions.
  85. Accuracy of k-nearest neighbor across different values of k;
  86. Accuracy of k-nearest neighbor for each kind of label;
  87. Accuracy of logistic regression for penalty values of 1 and 2;
  88. Accuracy of logistic regression for each kind of label;
  89. Accuracy of linear SVM across different kind of labels.
  90. Searchlight image of high arousal emotions vs low arousal emotions, and for distinct specific emotions.
  91. Accuracy of k-nearest neighbor across different values of k;
  92. Accuracy of k-nearest neighbor for each kind of label;
  93. Accuracy of logistic regression for penalty values of 1 and 2;
  94. Accuracy of logistic regression for each kind of label;
  95. Accuracy of linear SVM across different kind of labels.
  96. Searchlight image of high arousal emotions vs low arousal emotions, and for distinct specific emotions.
  97. Accuracy of k-nearest neighbor across different values of k;
  98. Accuracy of k-nearest neighbor for each kind of label;
  99. Accuracy of logistic regression for penalty values of 1 and 2;
  100. Accuracy of logistic regression for each kind of label;
  101. Accuracy of linear SVM across different kind of labels.
  102. Searchlight image of high arousal emotions vs low arousal emotions, and for distinct specific emotions.
  103. Accuracy of k-nearest neighbor across different values of k;
  104. Accuracy of k-nearest neighbor for each kind of label;
  105. Accuracy of logistic regression for penalty values of 1 and 2;
  106. Accuracy of logistic regression for each kind of label;
  107. Accuracy of linear SVM across different kind of labels.
  108. Searchlight image of high arousal emotions vs low arousal emotions, and for distinct specific emotions.
  109. Accuracy of k-nearest neighbor across different values of k;
  110. Accuracy of k-nearest neighbor for each kind of label;
  111. Accuracy of logistic regression for penalty values of 1 and 2;
  112. Accuracy of logistic regression for each kind of label;
  113. Accuracy of linear SVM across different kind of labels.
  114. Searchlight image of high arousal emotions vs low arousal emotions, and for distinct specific emotions.
  115. Accuracy of k-nearest neighbor across different values of k;
  116. Accuracy of k-nearest neighbor for each kind of label;
  117. Accuracy of logistic regression for penalty values of 1 and 2;
  118. Accuracy of logistic regression for each kind of label;
  119. Accuracy of linear SVM across different kind of labels.
  120. Searchlight image of high arousal emotions vs low arousal emotions, and for distinct specific emotions.
  121. Accuracy of k-nearest neighbor across different values of k;
  122. Accuracy of k-nearest neighbor for each kind of label;
  123. Accuracy of logistic regression for penalty values of 1 and 2;
  124. Accuracy of logistic regression for each kind of label;
  125. Accuracy of linear SVM across different kind of labels.
  126. Searchlight image of high arousal emotions vs low arousal emotions, and for distinct specific emotions.
  127. Accuracy of k-nearest neighbor across different values of k;
  128. Accuracy of k-nearest neighbor for each kind of label;
  129. Accuracy of logistic regression for penalty values of 1 and 2;
  130. Accuracy of logistic regression for each kind of label;
  131. Accuracy of linear SVM across different kind of labels.
  132. Searchlight image of high arousal emotions vs low arousal emotions, and for distinct specific emotions.
  133. Accuracy of k-nearest neighbor across different values of k;
  134. Accuracy of k-nearest neighbor for each kind of label;
  135. Accuracy of logistic regression for penalty values of 1 and 2;
  136. Accuracy of logistic regression for each kind of label;
  137. Accuracy of linear SVM across different kind of labels.
  138. Searchlight image of high arousal emotions vs low arousal emotions, and for distinct specific emotions.
  139. Accuracy of k-nearest neighbor across different values of k;
  140. Accuracy of k-nearest neighbor for each kind of label;
  141. Accuracy of logistic regression for penalty values of 1 and 2;
  142. Accuracy of logistic regression for each kind of label;
  143. Accuracy of linear SVM across different kind of labels.
  144. Searchlight image of high arousal emotions vs low arousal emotions, and for distinct specific emotions.
  145. Accuracy of k-nearest neighbor across different values of k;
  146. Accuracy of k-nearest neighbor for each kind of label;
  147. Accuracy of logistic regression for penalty values of 1 and 2;
  148. Accuracy of logistic regression for each kind of label;
  149. Accuracy of linear SVM across different kind of labels.
  150. Searchlight image of high arousal emotions vs low arousal emotions, and for distinct specific emotions.
  151. Accuracy of k-nearest neighbor across different values of k;
  152. Accuracy of k-nearest neighbor for each kind of label;
  153. Accuracy of logistic regression for penalty values of 1 and 2;
  154. Accuracy of logistic regression for each kind of label;
  155. Accuracy of linear SVM across different kind of labels.
  156. Searchlight image of high arousal emotions vs low arousal emotions, and for distinct specific emotions.
  157. Accuracy of k-nearest neighbor across different values of k;
  158. Accuracy of k-nearest neighbor for each kind of label;
  159. Accuracy of logistic regression for penalty values of 1 and 2;
  160. Accuracy of logistic regression for each kind of label;
  161. Accuracy of linear SVM across different kind of labels.
  162. Searchlight image of high arousal emotions vs low arousal emotions, and for distinct specific emotions.
  163. Accuracy of k-nearest neighbor across different values of k;
  164. Accuracy of k-nearest neighbor for each kind of label;
  165. Accuracy of logistic regression for penalty values of 1 and 2;
  166. Accuracy of logistic regression for each kind of label;
  167. Accuracy of linear SVM across different kind of labels.
  168. Searchlight image of high arousal emotions vs low arousal emotions, and for distinct specific emotions.
  169. Accuracy of k-nearest neighbor across different values of k;
  170. Accuracy of k-nearest neighbor for each kind of label;
  171. Accuracy of logistic regression for penalty values of 1 and 2;
  172. Accuracy of logistic regression for each kind of label;
  173. Accuracy of linear SVM across different kind of labels.
  174. Searchlight image of high arousal emotions vs low arousal emotions, and for distinct specific emotions.
  175. Accuracy of k-nearest neighbor across different values of k;
  176. Accuracy of k-nearest neighbor for each kind of label;
  177. Accuracy of logistic regression for penalty values of 1 and 2;
  178. Accuracy of logistic regression for each kind of label;
  179. Accuracy of linear SVM across different kind of labels.
  180. Searchlight image of high arousal emotions vs low arousal emotions, and for distinct specific emotions.
  181. Accuracy of k-nearest neighbor across different values of k;
  182. Accuracy of k-nearest neighbor for each kind of label;
  183. Accuracy of logistic regression for penalty values of 1 and 2;
  184. Accuracy of logistic regression for each kind of label;
  185. Accuracy of linear SVM across different kind of labels.
  186. Searchlight image of high arousal emotions vs low arousal emotions, and for distinct specific emotions.
  187. Accuracy of k-nearest neighbor across different values of k;
  188. Accuracy of k-nearest neighbor for each kind of label;
  189. Accuracy of logistic regression for penalty values of 1 and 2;
  190. Accuracy of logistic regression for each kind of label;
  191. Accuracy of linear SVM across different kind of labels.
  192. Searchlight image of high arousal emotions vs low arousal emotions, and for distinct specific emotions.
  193. Accuracy of k-nearest neighbor across different values of k;
  194. Accuracy of k-nearest neighbor for each kind of label;
  195. Accuracy of logistic regression for penalty values of 1 and 2;
  196. Accuracy of logistic regression for each kind of label;
  197. Accuracy of linear SVM across different kind of labels.
  198. Searchlight image of high arousal emotions vs low arousal emotions, and for distinct specific emotions.
  199. Accuracy of k-nearest neighbor across different values of k;
  200. Accuracy of k-nearest neighbor for each kind of label;
  201. Accuracy of logistic regression for penalty values of 1 and 2;
  202. Accuracy of logistic regression for each kind of label;
  203. Accuracy of linear SVM across different kind of labels.
  204. Searchlight image of high arousal emotions vs low arousal emotions, and for distinct specific emotions.
  205. Accuracy of k-nearest neighbor across different values of k;
  206. Accuracy of k-nearest neighbor for each kind of label;
  207. Accuracy of logistic regression for penalty values of 1 and 2;
  208. Accuracy of logistic regression for each kind of label;
  209. Accuracy of linear SVM across different kind of labels.
  210. Searchlight image of high arousal emotions vs low arousal emotions, and for distinct specific emotions.
  211. Accuracy of k-nearest neighbor across different values of k;
  212. Accuracy of k-nearest neighbor for each kind of label;
  213. Accuracy of logistic regression for penalty values of 1 and 2;
  214. Accuracy of logistic regression for each kind of label;
  215. Accuracy of linear SVM across different kind of labels.
  216. Searchlight image of high arousal emotions vs low arousal emotions, and for distinct specific emotions.
  217. Accuracy of k-nearest neighbor across different values of k;
  218. Accuracy of k-nearest neighbor for each kind of label;
  219. Accuracy of logistic regression for penalty values of 1 and 2;
  220. Accuracy of logistic regression for each kind of label;
  221. Accuracy of linear SVM across different kind of labels.
  222. Searchlight image of high arousal emotions vs low arousal emotions, and for distinct specific emotions.
  223. Accuracy of k-nearest neighbor across different values of k;
  224. Accuracy of k-nearest neighbor for each kind of label;
  225. Accuracy of logistic regression for penalty values of 1 and 2;
  226. Accuracy of logistic regression for each kind of label;
  227. Accuracy of linear SVM across different kind of labels.
  228. Searchlight image of high arousal emotions vs low arousal emotions, and for distinct specific emotions.
  229. Accuracy of k-nearest neighbor across different values of k;
  230. Accuracy of k-nearest neighbor for each kind of label;
  231. Accuracy of logistic regression for penalty values of 1 and 2;
  232. Accuracy of logistic regression for each kind of label;
  233. Accuracy of linear SVM across different kind of labels.
  234. Searchlight image of high arousal emotions vs low arousal emotions, and for distinct specific emotions.
  235. Accuracy of k-nearest neighbor across different values of k;
  236. Accuracy of k-nearest neighbor for each kind of label;
  237. Accuracy of logistic regression for penalty values of 1 and 2;
  238. Accuracy of logistic regression for each kind of label;
  239. Accuracy of linear SVM across different kind of labels.
  240. Searchlight image of high arousal emotions vs low arousal emotions, and for distinct specific emotions.
  241. Accuracy of k-nearest neighbor across different values of k;
  242. Accuracy of k-nearest neighbor for each kind of label;
  243. Accuracy of logistic regression for penalty values of 1 and 2;
  244. Accuracy of logistic regression for each kind of label;
  245. Accuracy of linear SVM across different kind of labels.
  246. Searchlight image of high arousal emotions vs low arousal emotions, and for distinct specific emotions.
  247. Accuracy of k-nearest neighbor across different values of k;
  248. Accuracy of k-nearest neighbor for each kind of label;
  249. Accuracy of logistic regression for penalty values of 1 and 2;
  250. Accuracy of logistic regression for each kind of label;
  251. Accuracy of linear SVM across different kind of labels.
  252. Searchlight image of high arousal emotions vs low arousal emotions, and for distinct specific emotions.
  253. Accuracy of k-nearest neighbor across different values of k;
  254. Accuracy of k-nearest neighbor for each kind of label;
  255. Accuracy of logistic regression for penalty values of 1 and 2;
  256. Accuracy of logistic regression for each kind of label;
  257. Accuracy of linear SVM across different kind of labels.
  258. Searchlight image of high arousal emotions vs low arousal emotions, and for distinct specific emotions.
  259. Accuracy of k-nearest neighbor across different values of k;
  260. Accuracy of k-nearest neighbor for each kind of label;
  261. Accuracy of logistic regression for penalty values of 1 and 2;
  262. Accuracy of logistic regression for each kind of label;
  263. Accuracy of linear SVM across different kind of labels.
  264. Searchlight image of high arousal emotions vs low arousal emotions, and for distinct specific emotions.
  265. Accuracy of k-nearest neighbor across different values of k;
  266. Accuracy of k-nearest neighbor for each kind of label;
  267. Accuracy of logistic regression for penalty values of 1 and 2;
  268. Accuracy of logistic regression for each kind of label;
  269. Accuracy of linear SVM across different kind of labels.
  270. Searchlight image of high arousal emotions vs low arousal emotions, and for distinct specific emotions.
  271. Accuracy of k-nearest neighbor across different values of k;
  272. Accuracy of k-nearest neighbor for each kind of label;
  273. Accuracy of logistic regression for penalty values of 1 and 2;
  274. Accuracy of logistic regression for each kind of label;
  275. Accuracy of linear SVM across different kind of labels.
  276. Searchlight image of high arousal emotions vs low arousal emotions, and for distinct specific emotions.
  277. Accuracy of k-nearest neighbor across different values of k;
  278. Accuracy of k-nearest neighbor for each kind of label;
  279. Accuracy of logistic regression for penalty values of 1 and 2;
  280. Accuracy of logistic regression for each kind of label;
  281. Accuracy of linear SVM across different kind of labels.
  282. Searchlight image of high arousal emotions vs low arousal emotions, and for distinct specific emotions.
  283. Accuracy of k-nearest neighbor across different values of k;
  284. Accuracy of k-nearest neighbor for each kind of label;
  285. Accuracy of logistic regression for penalty values of 1 and 2;
  286. Accuracy of logistic regression for each kind of label;
  287. Accuracy of linear SVM across different kind of labels.
  288. Searchlight image of high arousal emotions vs low arousal emotions, and for distinct specific emotions.
  289. Accuracy of k-nearest neighbor across different values of k;
  290. Accuracy of k-nearest neighbor for each kind of label;
  291. Accuracy of logistic regression for penalty values of 1 and 2;
  292. Accuracy of logistic regression for each kind of label;
  293. Accuracy of linear SVM across different kind of labels.
  294. Searchlight image of high arousal emotions vs low arousal emotions, and for distinct specific emotions.
  295. Accuracy of k-nearest neighbor across different values of k;
  296. Accuracy of k-nearest neighbor for each kind of label;
  297. Accuracy of logistic regression for penalty values of 1 and 2;
  298. Accuracy of logistic regression for each kind of label;
  299. Accuracy of linear SVM across different kind of labels.
  300. Searchlight image of high arousal emotions vs low arousal emotions, and for distinct specific emotions.
  301. Accuracy of k-nearest neighbor across different values of k;
  302. Accuracy of k-nearest neighbor for each kind of label;
  303. Accuracy of logistic regression for penalty values of 1 and 2;
  304. Accuracy of logistic regression for each kind of label;
  305. Accuracy of linear SVM across different kind of labels.
  306. Searchlight image of high arousal emotions vs low arousal emotions, and for distinct specific emotions.
  307. Accuracy of k-nearest neighbor across different values of k;
  308. Accuracy of k-nearest neighbor for each kind of label;
  309. Accuracy of logistic regression for penalty values of 1 and 2;
  310. Accuracy of logistic regression for each kind of label;
  311. Accuracy of linear SVM across different kind of labels.
  312. Searchlight image of high arousal emotions vs low arousal emotions, and for distinct specific emotions.
Four and five will be in our preliminary results.
The first five will describe which machine learning models work best for varying types of emotional labels. For each, we will need to:
  1. Clean the label data, and select the all sets of labels;
  2. Mask, smooth, and average fMRI data; apply each to their correct labels;
  3. Apply knn, SVM, and logistical regression to each dataset, for each parameter (ie, for each value of k)
  4. Gather predictive accuracy data for each run.
  5. Put predictive accuracy data in ggplot2.
  6. Clean the label data, and select the all sets of labels;
  7. Mask, smooth, and average fMRI data; apply each to their correct labels;
  8. Apply knn, SVM, and logistical regression to each dataset, for each parameter (ie, for each value of k)
  9. Gather predictive accuracy data for each run.
  10. Put predictive accuracy data in ggplot2.
  11. Clean the label data, and select the all sets of labels;
  12. Mask, smooth, and average fMRI data; apply each to their correct labels;
  13. Apply knn, SVM, and logistical regression to each dataset, for each parameter (ie, for each value of k)
  14. Gather predictive accuracy data for each run.
  15. Put predictive accuracy data in ggplot2.
  16. Clean the label data, and select the all sets of labels;
  17. Mask, smooth, and average fMRI data; apply each to their correct labels;
  18. Apply knn, SVM, and logistical regression to each dataset, for each parameter (ie, for each value of k)
  19. Gather predictive accuracy data for each run.
  20. Put predictive accuracy data in ggplot2.
  21. Clean the label data, and select the all sets of labels;
  22. Mask, smooth, and average fMRI data; apply each to their correct labels;
  23. Apply knn, SVM, and logistical regression to each dataset, for each parameter (ie, for each value of k)
  24. Gather predictive accuracy data for each run.
  25. Put predictive accuracy data in ggplot2.
  26. Clean the label data, and select the all sets of labels;
  27. Mask, smooth, and average fMRI data; apply each to their correct labels;
  28. Apply knn, SVM, and logistical regression to each dataset, for each parameter (ie, for each value of k)
  29. Gather predictive accuracy data for each run.
  30. Put predictive accuracy data in ggplot2.
  31. Clean the label data, and select the all sets of labels;
  32. Mask, smooth, and average fMRI data; apply each to their correct labels;
  33. Apply knn, SVM, and logistical regression to each dataset, for each parameter (ie, for each value of k)
  34. Gather predictive accuracy data for each run.
  35. Put predictive accuracy data in ggplot2.
  36. Clean the label data, and select the all sets of labels;
  37. Mask, smooth, and average fMRI data; apply each to their correct labels;
  38. Apply knn, SVM, and logistical regression to each dataset, for each parameter (ie, for each value of k)
  39. Gather predictive accuracy data for each run.
  40. Put predictive accuracy data in ggplot2.
  41. Clean the label data, and select the all sets of labels;
  42. Mask, smooth, and average fMRI data; apply each to their correct labels;
  43. Apply knn, SVM, and logistical regression to each dataset, for each parameter (ie, for each value of k)
  44. Gather predictive accuracy data for each run.
  45. Put predictive accuracy data in ggplot2.
  46. Clean the label data, and select the all sets of labels;
  47. Mask, smooth, and average fMRI data; apply each to their correct labels;
  48. Apply knn, SVM, and logistical regression to each dataset, for each parameter (ie, for each value of k)
  49. Gather predictive accuracy data for each run.
  50. Put predictive accuracy data in ggplot2.
  51. Clean the label data, and select the all sets of labels;
  52. Mask, smooth, and average fMRI data; apply each to their correct labels;
  53. Apply knn, SVM, and logistical regression to each dataset, for each parameter (ie, for each value of k)
  54. Gather predictive accuracy data for each run.
  55. Put predictive accuracy data in ggplot2.
  56. Clean the label data, and select the all sets of labels;
  57. Mask, smooth, and average fMRI data; apply each to their correct labels;
  58. Apply knn, SVM, and logistical regression to each dataset, for each parameter (ie, for each value of k)
  59. Gather predictive accuracy data for each run.
  60. Put predictive accuracy data in ggplot2.
  61. Clean the label data, and select the all sets of labels;
  62. Mask, smooth, and average fMRI data; apply each to their correct labels;
  63. Apply knn, SVM, and logistical regression to each dataset, for each parameter (ie, for each value of k)
  64. Gather predictive accuracy data for each run.
  65. Put predictive accuracy data in ggplot2.
  66. Clean the label data, and select the all sets of labels;
  67. Mask, smooth, and average fMRI data; apply each to their correct labels;
  68. Apply knn, SVM, and logistical regression to each dataset, for each parameter (ie, for each value of k)
  69. Gather predictive accuracy data for each run.
  70. Put predictive accuracy data in ggplot2.
  71. Clean the label data, and select the all sets of labels;
  72. Mask, smooth, and average fMRI data; apply each to their correct labels;
  73. Apply knn, SVM, and logistical regression to each dataset, for each parameter (ie, for each value of k)
  74. Gather predictive accuracy data for each run.
  75. Put predictive accuracy data in ggplot2.
  76. Clean the label data, and select the all sets of labels;
  77. Mask, smooth, and average fMRI data; apply each to their correct labels;
  78. Apply knn, SVM, and logistical regression to each dataset, for each parameter (ie, for each value of k)
  79. Gather predictive accuracy data for each run.
  80. Put predictive accuracy data in ggplot2.
  81. Clean the label data, and select the all sets of labels;
  82. Mask, smooth, and average fMRI data; apply each to their correct labels;
  83. Apply knn, SVM, and logistical regression to each dataset, for each parameter (ie, for each value of k)
  84. Gather predictive accuracy data for each run.
  85. Put predictive accuracy data in ggplot2.
  86. Clean the label data, and select the all sets of labels;
  87. Mask, smooth, and average fMRI data; apply each to their correct labels;
  88. Apply knn, SVM, and logistical regression to each dataset, for each parameter (ie, for each value of k)
  89. Gather predictive accuracy data for each run.
  90. Put predictive accuracy data in ggplot2.
  91. Clean the label data, and select the all sets of labels;
  92. Mask, smooth, and average fMRI data; apply each to their correct labels;
  93. Apply knn, SVM, and logistical regression to each dataset, for each parameter (ie, for each value of k)
  94. Gather predictive accuracy data for each run.
  95. Put predictive accuracy data in ggplot2.
  96. Clean the label data, and select the all sets of labels;
  97. Mask, smooth, and average fMRI data; apply each to their correct labels;
  98. Apply knn, SVM, and logistical regression to each dataset, for each parameter (ie, for each value of k)
  99. Gather predictive accuracy data for each run.
  100. Put predictive accuracy data in ggplot2.
  101. Clean the label data, and select the all sets of labels;
  102. Mask, smooth, and average fMRI data; apply each to their correct labels;
  103. Apply knn, SVM, and logistical regression to each dataset, for each parameter (ie, for each value of k)
  104. Gather predictive accuracy data for each run.
  105. Put predictive accuracy data in ggplot2.
  106. Clean the label data, and select the all sets of labels;
  107. Mask, smooth, and average fMRI data; apply each to their correct labels;
  108. Apply knn, SVM, and logistical regression to each dataset, for each parameter (ie, for each value of k)
  109. Gather predictive accuracy data for each run.
  110. Put predictive accuracy data in ggplot2.
  111. Clean the label data, and select the all sets of labels;
  112. Mask, smooth, and average fMRI data; apply each to their correct labels;
  113. Apply knn, SVM, and logistical regression to each dataset, for each parameter (ie, for each value of k)
  114. Gather predictive accuracy data for each run.
  115. Put predictive accuracy data in ggplot2.
  116. Clean the label data, and select the all sets of labels;
  117. Mask, smooth, and average fMRI data; apply each to their correct labels;
  118. Apply knn, SVM, and logistical regression to each dataset, for each parameter (ie, for each value of k)
  119. Gather predictive accuracy data for each run.
  120. Put predictive accuracy data in ggplot2.
  121. Clean the label data, and select the all sets of labels;
  122. Mask, smooth, and average fMRI data; apply each to their correct labels;
  123. Apply knn, SVM, and logistical regression to each dataset, for each parameter (ie, for each value of k)
  124. Gather predictive accuracy data for each run.
  125. Put predictive accuracy data in ggplot2.
  126. Clean the label data, and select the all sets of labels;
  127. Mask, smooth, and average fMRI data; apply each to their correct labels;
  128. Apply knn, SVM, and logistical regression to each dataset, for each parameter (ie, for each value of k)
  129. Gather predictive accuracy data for each run.
  130. Put predictive accuracy data in ggplot2.
  131. Clean the label data, and select the all sets of labels;
  132. Mask, smooth, and average fMRI data; apply each to their correct labels;
  133. Apply knn, SVM, and logistical regression to each dataset, for each parameter (ie, for each value of k)
  134. Gather predictive accuracy data for each run.
  135. Put predictive accuracy data in ggplot2.
  136. Clean the label data, and select the all sets of labels;
  137. Mask, smooth, and average fMRI data; apply each to their correct labels;
  138. Apply knn, SVM, and logistical regression to each dataset, for each parameter (ie, for each value of k)
  139. Gather predictive accuracy data for each run.
  140. Put predictive accuracy data in ggplot2.
  141. Clean the label data, and select the all sets of labels;
  142. Mask, smooth, and average fMRI data; apply each to their correct labels;
  143. Apply knn, SVM, and logistical regression to each dataset, for each parameter (ie, for each value of k)
  144. Gather predictive accuracy data for each run.
  145. Put predictive accuracy data in ggplot2.
  146. Clean the label data, and select the all sets of labels;
  147. Mask, smooth, and average fMRI data; apply each to their correct labels;
  148. Apply knn, SVM, and logistical regression to each dataset, for each parameter (ie, for each value of k)
  149. Gather predictive accuracy data for each run.
  150. Put predictive accuracy data in ggplot2.
  151. Clean the label data, and select the all sets of labels;
  152. Mask, smooth, and average fMRI data; apply each to their correct labels;
  153. Apply knn, SVM, and logistical regression to each dataset, for each parameter (ie, for each value of k)
  154. Gather predictive accuracy data for each run.
  155. Put predictive accuracy data in ggplot2.
  156. Clean the label data, and select the all sets of labels;
  157. Mask, smooth, and average fMRI data; apply each to their correct labels;
  158. Apply knn, SVM, and logistical regression to each dataset, for each parameter (ie, for each value of k)
  159. Gather predictive accuracy data for each run.
  160. Put predictive accuracy data in ggplot2.
  161. Clean the label data, and select the all sets of labels;
  162. Mask, smooth, and average fMRI data; apply each to their correct labels;
  163. Apply knn, SVM, and logistical regression to each dataset, for each parameter (ie, for each value of k)
  164. Gather predictive accuracy data for each run.
  165. Put predictive accuracy data in ggplot2.
  166. Clean the label data, and select the all sets of labels;
  167. Mask, smooth, and average fMRI data; apply each to their correct labels;
  168. Apply knn, SVM, and logistical regression to each dataset, for each parameter (ie, for each value of k)
  169. Gather predictive accuracy data for each run.
  170. Put predictive accuracy data in ggplot2.
  171. Clean the label data, and select the all sets of labels;
  172. Mask, smooth, and average fMRI data; apply each to their correct labels;
  173. Apply knn, SVM, and logistical regression to each dataset, for each parameter (ie, for each value of k)
  174. Gather predictive accuracy data for each run.
  175. Put predictive accuracy data in ggplot2.
  176. Clean the label data, and select the all sets of labels;
  177. Mask, smooth, and average fMRI data; apply each to their correct labels;
  178. Apply knn, SVM, and logistical regression to each dataset, for each parameter (ie, for each value of k)
  179. Gather predictive accuracy data for each run.
  180. Put predictive accuracy data in ggplot2.
  181. Clean the label data, and select the all sets of labels;
  182. Mask, smooth, and average fMRI data; apply each to their correct labels;
  183. Apply knn, SVM, and logistical regression to each dataset, for each parameter (ie, for each value of k)
  184. Gather predictive accuracy data for each run.
  185. Put predictive accuracy data in ggplot2.
  186. Clean the label data, and select the all sets of labels;
  187. Mask, smooth, and average fMRI data; apply each to their correct labels;
  188. Apply knn, SVM, and logistical regression to each dataset, for each parameter (ie, for each value of k)
  189. Gather predictive accuracy data for each run.
  190. Put predictive accuracy data in ggplot2.
  191. Clean the label data, and select the all sets of labels;
  192. Mask, smooth, and average fMRI data; apply each to their correct labels;
  193. Apply knn, SVM, and logistical regression to each dataset, for each parameter (ie, for each value of k)
  194. Gather predictive accuracy data for each run.
  195. Put predictive accuracy data in ggplot2.
  196. Clean the label data, and select the all sets of labels;
  197. Mask, smooth, and average fMRI data; apply each to their correct labels;
  198. Apply knn, SVM, and logistical regression to each dataset, for each parameter (ie, for each value of k)
  199. Gather predictive accuracy data for each run.
  200. Put predictive accuracy data in ggplot2.
  201. Clean the label data, and select the all sets of labels;
  202. Mask, smooth, and average fMRI data; apply each to their correct labels;
  203. Apply knn, SVM, and logistical regression to each dataset, for each parameter (ie, for each value of k)
  204. Gather predictive accuracy data for each run.
  205. Put predictive accuracy data in ggplot2.
  206. Clean the label data, and select the all sets of labels;
  207. Mask, smooth, and average fMRI data; apply each to their correct labels;
  208. Apply knn, SVM, and logistical regression to each dataset, for each parameter (ie, for each value of k)
  209. Gather predictive accuracy data for each run.
  210. Put predictive accuracy data in ggplot2.
  211. Clean the label data, and select the all sets of labels;
  212. Mask, smooth, and average fMRI data; apply each to their correct labels;
  213. Apply knn, SVM, and logistical regression to each dataset, for each parameter (ie, for each value of k)
  214. Gather predictive accuracy data for each run.
  215. Put predictive accuracy data in ggplot2.
  216. Clean the label data, and select the all sets of labels;
  217. Mask, smooth, and average fMRI data; apply each to their correct labels;
  218. Apply knn, SVM, and logistical regression to each dataset, for each parameter (ie, for each value of k)
  219. Gather predictive accuracy data for each run.
  220. Put predictive accuracy data in ggplot2.
  221. Clean the label data, and select the all sets of labels;
  222. Mask, smooth, and average fMRI data; apply each to their correct labels;
  223. Apply knn, SVM, and logistical regression to each dataset, for each parameter (ie, for each value of k)
  224. Gather predictive accuracy data for each run.
  225. Put predictive accuracy data in ggplot2.
  226. Clean the label data, and select the all sets of labels;
  227. Mask, smooth, and average fMRI data; apply each to their correct labels;
  228. Apply knn, SVM, and logistical regression to each dataset, for each parameter (ie, for each value of k)
  229. Gather predictive accuracy data for each run.
  230. Put predictive accuracy data in ggplot2.
  231. Clean the label data, and select the all sets of labels;
  232. Mask, smooth, and average fMRI data; apply each to their correct labels;
  233. Apply knn, SVM, and logistical regression to each dataset, for each parameter (ie, for each value of k)
  234. Gather predictive accuracy data for each run.
  235. Put predictive accuracy data in ggplot2.
  236. Clean the label data, and select the all sets of labels;
  237. Mask, smooth, and average fMRI data; apply each to their correct labels;
  238. Put predictive accuracy data in ggplot2.
  239. Clean the label data, and select the all sets of labels;
  240. Mask, smooth, and average fMRI data; apply each to their correct labels;
  241. Apply knn, SVM, and logistical regression to each dataset, for each parameter (ie, for each value of k)
  242. Gather predictive accuracy data for each run.
  243. Put predictive accuracy data in ggplot2.
  244. Clean the label data, and select the all sets of labels;
  245. Mask, smooth, and average fMRI data; apply each to their correct labels;
  246. Apply knn, SVM, and logistical regression to each dataset, for each parameter (ie, for each value of k)
  247. Gather predictive accuracy data for each run.
  248. Put predictive accuracy data in ggplot2.
  249. Clean the label data, and select the all sets of labels;
  250. Mask, smooth, and average fMRI data; apply each to their correct labels;
  251. Apply knn, SVM, and logistical regression to each dataset, for each parameter (ie, for each value of k)
  252. Gather predictive accuracy data for each run.
  253. Put predictive accuracy data in ggplot2.
  254. Clean the label data, and select the all sets of labels;
  255. Mask, smooth, and average fMRI data; apply each to their correct labels;
  256. Apply knn, SVM, and logistical regression to each dataset, for each parameter (ie, for each value of k)
  257. Gather predictive accuracy data for each run.
  258. Put predictive accuracy data in ggplot2.
The final image will include the first three, and i