Rajeev Raizada (Cornell Univ.) "The brain's linguistic representations: from neural population codes to syllables and semantics"
Mon, January 23, 2012 • 3:00 PM - 4:30 PM • UTC 3.134
In seeking to understand how the human brain represents language, neuroimaging faces a serious problem: brain images do not reveal mental representations. Recent work in neural decoding has shown that classifier algorithms can recover information from individual subjects' fMRI data about the tasks and stimuli that evoked that activation. However, that decoding does not in itself tell us whether the classifiers extract information that is actually used by the brain. Nor does it tell us how the brain represents that information.
Addressing the first of those questions, I will present evidence from speech perception showing that multivoxel fMRI patterns can predict individual differences in people's behavioral ability to discriminate between heard syllables. Addressing the second, I will show how the structure of linguistic representations can be related to the structure of similarity relations between neural activation patterns.
This generates a novel hypothesis about how the brain represents the meanings of words, which can be stated very simply: neural similarity matches semantic similarity. The predictive power of this approach can be demonstrated by decoding the meanings of words from fMRI activation: it achieves substantially greater accuracy than any previously published method. Finally, I will discuss some of the challenges and new questions raised by this research.