Guest 4

Lecture 16 - February 10th, 2017

Guest Lecture

By Alona Fyshe

My research question:

How doees the brain encode (represent) the meaning of words? How does it combine word meaning to represent more complex concepts?

Siri and the Long tail:

Image reference

Outline

  • Brain imaging
  • Word meaning in the brain
  • Words in context

Brain imaging technology:

  • Functional Magnetic Resonance Imaging (fMRI)
    • Measuers changes in blood oxygenation (BOLD)
      • Good spatial resolution
      • Poor time resolution
  • MEG
    • Magnetoecephalography (magnet / brain / graph)
    • Magneted sheeted room -> because earth
      • 102 sensor position
      • 3 sensors / position
      • Poorer spatial resolution than fMRI
      • Great time resolution
      • Great for language studies

Meaning in the brain:

  • What would it mean to have a model of meaning representation in the brain?
    • Which brain areas acKvate for parKcular words
    • Predict the word a person is reading from their brain acKvaKon

Image reference 1 Image reference 2

How to build a model?

Show a computer algorithm many examples.
Training
Find Patterns
Test

Predict unseen words
Find Patterns

You can't build model for every word
    Solution: Word Features (eg. edible, alive)

Word prediction via features

Ex. The Brain Data

  • 9 subjects fMRI , 9 subjecy MEG
  • 60 concrete noun w/ pictures

Word Features

Co-occurence with verbs: (MRI)

  • Accuracy: 77%

Dog Example

20 Questions: (MEG)

  • Array of numbers = Identifier of the Word

Word Array

  • Visual Features (Of the Word and the Image)
    • Posterior Regions
      • Detected in 150ms-250ms
  • Word Meaning
    • Distrubuted throughout the brain
      • Detected in 200-600ms