Detection and Localisation of Neural Responses to Linguistic - PowerPoint PPT Presentation

Detection and Localisation of Neural Responses to Linguistic Phenomena using Machine Learning Student: Mehdi Parviz Supervisor: Mark Johnson Department of Computing Macquarie University COMP901, 2012 1 / 23

Outline Introduction Background Magnetoencephalography N400 Response Current Analysing Methods Approach Multivariate Analysis Logistic Regression Results Experimental Data Results 2 / 23

Outline Introduction Background Magnetoencephalography N400 Response Current Analysing Methods Approach Multivariate Analysis Logistic Regression Results Experimental Data Results 3 / 23

Background ◮ Main challenge ◮ localisation and detection of neuronal activity ◮ Non-invasive functional brain imaging ◮ Huge amount of data ◮ High dimensional space ◮ Noisy ◮ Very complex ◮ Analysing methods ◮ Univariate ◮ Multivariate 4 / 23

Outline Introduction Background Magnetoencephalography N400 Response Current Analysing Methods Approach Multivariate Analysis Logistic Regression Results Experimental Data Results 5 / 23

Magnetoencephalography (MEG) ◮ Non-invasive technique for imaging brain activity ◮ Higher temporal resolution than fMRI (1 ms vs. 1-4 s) ◮ Higher spatial resolution than EEG (15 mm vs. 20 mm) Figure: MEG Machine 6 / 23

Outline Introduction Background Magnetoencephalography N400 Response Current Analysing Methods Approach Multivariate Analysis Logistic Regression Results Experimental Data Results 7 / 23

N400 Response ◮ Event related potentials (ERPs) ◮ Synchronized activation of neuronal networks ◮ Generated by external stimuli ◮ N400 response ◮ Broad negative deflection of the ERPs ◮ Peaks 400 ms after post-stimulus onset ◮ N400 occurs in sentences containing semantically unexpected or anomalous words ◮ A sparrow is a kind of building ◮ A sparrow is a kind of bird 8 / 23

Outline Introduction Background Magnetoencephalography N400 Response Current Analysing Methods Approach Multivariate Analysis Logistic Regression Results Experimental Data Results 9 / 23

Current Analysing Methods ◮ Using univariate analysis methods ◮ Finding some predefined ERPs ◮ P600: a sentence with singular subject and plural verb ◮ Computing statistical significance of the brain response to the stimuli 10 / 23

Drawbacks of univariate methods ◮ Averaging over a large number of trials to increase the signal-to-noise ◮ Removing information which are not time locked ◮ Assuming independence between variables ◮ Ignoring potential covariance between neighboring or distant units ◮ Searching for highly localised response ◮ Ignoring several sources of activity ◮ Removing causal structure of the brain response 11 / 23

Outline Introduction Background Magnetoencephalography N400 Response Current Analysing Methods Approach Multivariate Analysis Logistic Regression Results Experimental Data Results 12 / 23

Multivariate Analysis ◮ Whole set of variables are analysed together ◮ Dependency between variables are considered ◮ Multiple source and causal structure can be addressed Figure: MEG Machine 13 / 23

Outline Introduction Background Magnetoencephalography N400 Response Current Analysing Methods Approach Multivariate Analysis Logistic Regression Results Experimental Data Results 14 / 23

Logistic Regression ◮ Conditional probability � log Pr ( Setence Type | Brain Respose ) = w j × x j (1) ◮ Maximising the likelihood � max log Pr ( Setence Type | Brain Respose ) (2) W ◮ Sparse solution using l1-norm regularisation 15 / 23

Outline Introduction Background Magnetoencephalography N400 Response Current Analysing Methods Approach Multivariate Analysis Logistic Regression Results Experimental Data Results 16 / 23

Experimental Data Sentences ◮ Stimuli consisted of 180 sentences drawn from the list published by Kalikow et al. (1997) ◮ 90 examples of “constraining context” sentences, i.e., with predictable endings (e.g. He got drunk in the local bar ) ◮ 90 examples of “non-constraining context” sentences, i.e., with unpredictable endings (e.g. He hopes Tom asked about the bar ) ◮ Each target word appears both in a constraining context sentence and in a non-constraining context sentence ◮ 10 catch trials consisting of sentences containing the word mouse ◮ 16 listeners 17 / 23

Experimental Data MEG Data MEG data were ◮ Extracted from 160 sensors and mapped into the brain space using SPM8 package ◮ Digitised with a sample rate of 1000 Hz ◮ Filtered with a bandpass of 0.1 to 40 Hz 18 / 23

Outline Introduction Background Magnetoencephalography N400 Response Current Analysing Methods Approach Multivariate Analysis Logistic Regression Results Experimental Data Results 19 / 23

Results Integrating over variables Time-Channel Time Channel Without integration 54.07 62.80 58.04 62.82 Table: The classification accuracy for each subject using all the data points (without integrating) and different types of integrating 20 / 23

Results Temporal Information Figure: Accuracy of detecting Context using different time windows 21 / 23

Results Brain Activity Figure: Feature weights for different temporal window 22 / 23

Summary ◮ Multivariate methods can detect and localise the brain response to linguistic inputs more accurately ◮ L1-norm regularisation can be applied to find sparse solution ◮ Brain response to the Context starts earlier than 400ms 23 / 23