Using Data to Understand the Brain
- 1. Using Data toUnderstand the Brain John Myles White March 26, 2012 1 / 57
- 2. We want tolearn how the brain works 2 / 57
- 3.
- 4.
- 5. How can learnabout the mind and brain using data and models? 5 / 57
- 6.
- 7.
- 8.
- 9. Brain damage victimstaught us about functional localization 9 / 57
- 10.
- 11. How can welearn more without shooting people? 11 / 57
- 12.
- 13. MRI can giveus: High-resolution anatomical images Real-time measurements of blood flow (fMRI) 13 / 57
- 14. A typical MRIexperimental data set contains: 10 MB of anatomical data ∼ 1 GB of fMRI data Time series of blood flow sampled every 2s Sampled at 60 x 60 x 60 voxels 14 / 57
- 15.
- 16.
- 17.
- 18. How can weuse blood flow measurements to learn about function? 18 / 57
- 19. Design an experimentthat randomly switches between tasks: Tapping one finger vs. sitting motionless Looking at faces vs. looking at places Thinking about people vs. thinking about objects 19 / 57
- 20.
- 21. Search blood flowdata for brain regions that respond to tasks 21 / 57
- 22. How can weanalyze blood flow data to perform this search? 22 / 57
- 23. The typical approachlooks at each voxel separately: Try to predict blood flow using task events Uses standard linear regression Many connections with signal processing 23 / 57
- 24. A newer approach,called MVPA, works in the reverse direction 24 / 57
- 25. In MVPA, wetry to predict tasks using blood flow 25 / 57
- 26.
- 27. We have fewertask-tagged brain images than voxels, so n < p 27 / 57
- 28. Need to useregularization to perform any analysis 28 / 57
- 29. Consider solving y= Xβ 29 / 57
- 30. When n =p, there may be a unique solution When n > p, we must choose an approximation When n < p, there are infinitely many exact solutions To find “correct” solution, we have to introduce constraints 30 / 57
- 31. In linear regression,we minimize (y − Xβ)2 31 / 57
- 32. In L2 regularization,we minimize (y − Xβ)2 + λβ 2 32 / 57
- 33. L2 constraint letsus solve n < p case 33 / 57
- 34. But L2 oftendoes not help us find the few voxels that matter most 34 / 57
- 35. We can tryanother regularization system called L1 regularization 35 / 57
- 36. In L1 regularization,we minimize (y − Xβ)2 + λ|β| 36 / 57
- 37. L1 regularization isvery modern Objective function is not differentiable But is convex and can be minimized computationally Solution, β ∗ , to minimization problem is typically sparse 37 / 57
- 38. A sparse solutionis one in which most features have a weight of 0 38 / 57
- 39. For MVPA, L1sparse solutions are, sadly, too sparse 39 / 57
- 40.
- 41. From neurology, weknow we should expect large clusters of voxels 41 / 57
- 42.
- 43. Best current approach:mix L1 and L2 43 / 57
- 44. In the ElasticNet, we minimize (y − Xβ)2 + λ1 |β| + λ2 β 2 44 / 57
- 45.
- 46. Improving localization usingMVPA is an ongoing problem 46 / 57
- 47. But MVPA hasalready changed our understanding of the brain 47 / 57
- 48. Forget localization. Focuson prediction 48 / 57
- 49. Imagine we havebuilt a classifier that identifies tasks correctly 49 / 57
- 50. We can usethe classifier to test how people think about other tasks 50 / 57
- 51. The free-recall task: Experimental subjects memorize items from 3 lists: 1. Locations 2. Faces 3. Objects Subjects then try to recall as many items as they can 51 / 57
- 52. Our theory: To remember items, you return to the mental state you were in when you memorized the lists Before you name any specific item, you return to the state concerned with that item’s category Only then can you name any specific items 52 / 57
- 53. Our approach: We train classifier to identify type of list being memorized We use classifier to assess mental state during free-recall 53 / 57
- 54.
- 55. In short, we’realready able to read minds 55 / 57
- 56. The real questionsfor the field are: How can we do it better? How can we learn more about the brain from fMRI data? 56 / 57
- 57. Any questions? 57 / 57