The study used fMRI and multi-voxel pattern analysis to examine the neural representations of semantic concepts in bilingual participants. The researchers found that abstract semantic representations of words like "horse" could be decoded from brain activity patterns in anterior temporal lobe regions, independently of the language used. This provides evidence that some ATL regions construct language-independent semantic representations and act as a semantic hub in the brain. The study demonstrates the ability of MVPA to examine fine-grained semantic representations across languages with higher sensitivity than traditional univariate analysis.
TataKelola dan KamSiber Kecerdasan Buatan v022.pdf
Brain-Based Translation Reveals Language-Independent Semantic Representations in ATL
1. Brain-Based Translation:
fMRI Decoding of Spoken Words
in Bilinguals Reveals Language-
Independent
Semantic Representations
in Anterior Temporal Lobe
Correia et al (2014). The Journal of Neuroscience
2. Statement of the
problem
What are the neural mechanisms
underlying the representation of
language-independent semantic concepts
in the brain? “Horse” in
English
“Paard” in
Dutch
3. General findings
Several regions in the ATL (Anterior
Temporal Lobe) seem to be responsible for
organizing language-independent semantic
concepts. Specifically, at the fine-grained
level of within-semantic-category
discriminations (e.g “animal” category).
“Horse” “Paard”
4. Claims Evidence
Patients with ATL brain lesions demonstrate
selective deficits of semantic knowledge, also
known as semantic dementia.
A crucial function of the ATL is to act
as a “semantic hub.” That is, language-
independent representations of words
and concepts are constructed in some
regions of the ATL.
Previous work
(Damasio et al., 1996; Pattersonet al., 2007)
5. The knowledge gap
in the literature
How this paper
attempts to address it
Using bilingual participants, fMRI, a unique
experimental paradigm, and a whole lot of
complicated data analyses that I hope we can
figure out together.
Lesion correlation evidence is insightful, but by no means
abundant or consistent across patients (e.g. what counts as a
lesion to ATL, what counts as semantic dementia).
Ideally, we’d want converging evidence from several kinds of
data (behavioral, lesion, neurolinguistic...etc.). Some work has
been done in this regard, but there exist some challenges to find
reliable ATL activation in neuroimaging. These include
“susceptibility artifacts” or “limited field of view.”
Recent studies have devised creative new techniques to
overcome these challenges. This paper is one of them.
6. Important
terms and
concepts
● Within-language discrimination
● Across-language generalization
● Block experiment design
● Univariate Analysis
● Multi-Voxel Pattern Analysis
● Cortex-based alignment
● Repetition Time (RT)?
● Acquisition Time (AT)?
● Baseline?
● Classification
● Training algorithms
● Searchlight method
● Support vector machine (SVM)
7. Experimental design
In separate Dutch and English blocks, bilingual participants were
asked to listen to individual animal nouns (e.g. horse) and to
detect occasional non-animal target nouns (e.g. bike). The goal of
the task was to help maintain a constant attention level
throughout the experiment to promote speech comprehension at
every word presentation.
Importantly, each word was spoken by 3 different female speakers,
which allowed for speaker invariant word discrimination.
8. Methodology
Participant criteria Ten bilinguals university
students. All L1 Dutch, proficient L2 English (3 males, 7
females).
Why bilinguals? Allows us to tease apart language-
specific & language-independent representations in a way
that monolingual data can’t.
Stimuli Consisted of Dutch and English spoken words
representing 4 different animals (bull, duck, horse, shark)
and 3 inanimate object words (bike, dress, suit)
Task Listen & detect non-animal words.
Data collection fMRI
Data analysis It’s complicated
9. Data analysis
1. Functional and anatomical image acquisition
2. fMRI data preprocessing
3. MRI data analysis univariate statistics
4. fMRI data analysis MVPA
5. Classification (Feature extraction, Feature selection, Cross-
validation)
6. Discriminative maps
10. Data analysis
1. Functional and anatomical image acquisition
2. fMRI data preprocessing
3. MRI data analysis univariate statistics
4. fMRI data analysis MVPA
5. Classification (Feature extraction, Feature selection, Cross-
validation)
6. Discriminative maps
Whole-brain analysis was achieved
with high-resolution (voxel size, 1x1x1
mm3) anatomical images.
11. Data analysis
1. Functional and anatomical image acquisition
2. fMRI data preprocessing
3. MRI data analysis univariate statistics
4. fMRI data analysis MVPA
5. Classification (Feature extraction, Feature selection, Cross-
validation)
6. Discriminative maps
They obtained an anatomically-aligned,
group-averaged 3D surface representation
so that subject data could be compared.
They did this by preprocessing and analyzing fMRI data
using Brain Voyager and custom-made MATLAB
routines.
● Functional data were corrected for (3D motion,
slice scan time differences, removing noise)
● Anatomical data were corrected for (intensity
inhomogeneity, transformed into Talairach
space).
● Functional & anatomical data then aligned to
create 4D volume time courses (cortex-based
alignment).
12. Data analysis
1. Functional and anatomical image acquisition
2. fMRI data preprocessing
3. fMRI data analysis univariate statistics
4. fMRI data analysis MVPA
5. Classification (Feature extraction, Feature selection, Cross-
validation)
6. Discriminative maps
Functional contrast maps were
constructed in order to identify the cortical regions
involved in the processing of spoken words.
This was done by:
(1) comparing activation to all animal words vs. baseline
across subjects
(2) combining all possible binary contrasts within nouns
of the same language
(3) grouping all equivalent nouns into single concepts
and contrasting all possible binary combinations of
concepts.
Functional contrast maps were
submitted to a whole-brain correction
criterion based on the estimate of the spatial
smoothness of the map.
Univariate analysis is the simplest form of
data analysis. It means that your analysis is
considering only one variable. It doesn’t
deal with causes or relationships (like
MVPA does). Its major purpose is to
describe, summarize and find patterns in
the data with respect to that one variable.
13. fMRI data Univariate analysis
Here we’re looking at the
fMRI responses elicited by
all animal words across
subjects, as computed by a
functional contrast
map (t statistics) comparing
all animal words versus
baseline.
14. Data analysis
1. Functional and anatomical image acquisition
2. fMRI data preprocessing
3. fMRI data analysis univariate statistics
4. fMRI data analysis MVPA
5. Classification (Feature extraction, Feature selection, Cross-
validation)
6. Discriminative maps
What can Multi-voxel pattern analysis
(MVPA) do that Univariate analysis can’t?
Detect differences between conditions with higher sensitivity than
conventional univariate analysis by focusing on the analysis and
comparison of distributed patterns of activity. Data from individual
voxels within a region are jointly analyzed.
Furthermore, MVPA is often presented in the context of "brain
reading" applications reporting that specific mental states or
representational content can be decoded from fMRI activity patterns
after performing a "training" or "learning phase. In this context,
MVPA tools are often referred to as classifiers or, more
generally, learning machines. The latter names stress that many
MVPA tools originate from a field called machine learning, a branch
of artificial intelligence.
BrainVoyager introduces a comprehensive set of MVPA tools for
locally distributed as well as more extended (sparse) patterns of
activation. The tools include a multivariate searchlight mapping
approach, which is used both for analyzing patterns in ROIs and for
discriminating patterns that are potentially spread out across the
whole brain.
BrainVoyager QX v2.8 | 2014 Rainer Goebel.
15. fMRI data analysis MVPA
Here we’re looking at the
statistical maps of
searchlight selections for
which the word
discrimination and the
language generalization
analyses yielded
accuracies significantly
above chance level
(50%).
16. Data analysis
1. Functional and anatomical image acquisition
2. fMRI data preprocessing
3. fMRI data analysis univariate statistics
4. fMRI data analysis MVPA
5. Classification (Feature extraction, Feature selection,
Cross-validation)
6. Discriminative maps
?
17. Data analysis
1. Functional and anatomical image acquisition
2. fMRI data preprocessing
3. fMRI data analysis univariate statistics
4. fMRI data analysis MVPA
5. Classification (Feature extraction, Feature selection, Cross-
validation)
6. Discriminative maps
Accuracy maps of within-language word
discrimination and across-language
word generalization were constructed.
Accuracy maps were averaged within each subject across
binary comparisons and cross-validation folds.
Thereafter, individual averaged accuracy maps were
projected onto the group-averaged cortical surface and
anatomically aligned using cortex-based alignment.
In the within-language discrimination
analysis, one map was produced per language and
subsequently combined into a single map by means of
conjunction analysis. The resulting discrimination map
thus depicts regions with consistent sensitivity in English
and Dutch.
For the across-language word generalization, one map
was produced from all possible binary language
generalizations.
20. Conclusions
● That abstract horse representation exists in
the brain and can be located. Brain-based decoding
of individual spoken words at the fine-grained level of within-
semantic category is possible within and across the first and
second languages of bilingual adults.
● That representation exists somewhere in the
ATL. Specifically, localized clusters in the left anterior
temporal lobe, the left angular gyrus and the posterior bank
of the left postcentral gyrus, the right posterior superior
temporal sulcus/superior temporal gyrus, the right medial
anterior temporal lobe, the right anterior insula, and bilateral
occipital cortex.
● MVPA is a good approach for this kind of research
question. Results indicate the benefits of MVPA based on
the generalization of the pattern information across specific
stimulus dimensions. This approach enabled examining the
representation of spoken words independently of the
speaker and the representation of semantic– conceptual
information independently of the input language.
21. Questions
● So what? Why should we care where
representations are or where processes
happen in the brain?
● Could we translate this study (an fMRI
study examining the location of abstract
semantic representations in the brain) with
other neurolinguistic tools (MEG,
EEG...etc.). If so, how? Benefits or
advantages of doing that?
● Do you feel their experimental design
tested what they thought they were
testing?
● If you were to change something about
their experimental design or data analysis,
what would it be?
Editor's Notes
Does that abstract representation exist? If so, where? How can we find out?
In this approach, each voxel is visited, as in a standard univariate analysis, but instead of using only data of the visited voxel for analysis, several voxels in the neighborhood are included forming a set of features for joined multivariate analysis.
Support Vector Machines (SVM) are supervised learning models with associated learning algorithms that analyze data used for classification and regression analysis. Given a set of training examples, each marked as belonging to one or the other of two categories, an SVM training algorithm builds a model that assigns new examples to one category or the other,
The difference between the Acquisition Time (TA) and
Repetition Time (TR) introduced a silent gap used for the presentation of the auditory stimuli.
unctional data were 3D motion corrected
(trilinear interpolation), corrected for slice scan time differences,and were temporally filtered by removing frequency components of five or less cycles per time course.
Anatomical data were corrected for intensity inhomogeneity and transformed into Talairach space.
Functional data were then aligned with the anatomical data and transformed into the same space to
create 4D volume time courses.
Individual cortical surfaces were reconstructed from gray–white matter segmentations of the anatomical data and aligned using a moving target-group average approach based on
curvature information (cortex-based alignment)
Univariate effects were analyzed using a random-effects general linear model (GLM). A single predictor per stimulus condition was convoluted with a double gamma hemodynamic response function.
Univariate effects were analyzed using a random-effects general linear model (GLM). A single predictor per stimulus condition was convoluted with a double gamma hemodynamic response function.
The idea of this step was to investigate speech content information in the fMRI responses. To do this, they used a supervised machine learning algorithm combined with single-trial multivoxel classification. Classifications were performed to evaluate whether patterns of voxels conveyed information on the identity of spoken words (within-language word discrimination) and their
language-invariant representations (across-language word generalization).
Within-language word discrimination entailed training a classifier to discriminate between two words of the same language (e.g., horse vs duck) and testing in the same words spoken by a speaker not included in
the training phase. Across-language word generalization was performed by training a classifier to discriminate between two words within one language (e.g., horse vs duck) and testing in the translational equivalent
words within the other language (e.g., paard vs eend), thus relying on language-independent information of spoken words representing equivalent
concepts in Dutch and English (Fig. 1A).
Univariate effects were analyzed using a random-effects general linear model (GLM). A single predictor per stimulus condition was convoluted with a double gamma hemodynamic response function.
The idea of this step was to investigate speech content information in the fMRI responses. To do this, they used a supervised machine learning algorithm combined with single-trial multivoxel classification. Classifications were performed to evaluate whether patterns of voxels conveyed information on the identity of spoken words (within-language word discrimination) and their
language-invariant representations (across-language word generalization).
Within-language word discrimination entailed training a classifier to discriminate between two words of the same language (e.g., horse vs duck) and testing in the same words spoken by a speaker not included in
the training phase. Across-language word generalization was performed by training a classifier to discriminate between two words within one language (e.g., horse vs duck) and testing in the translational equivalent
words within the other language (e.g., paard vs eend), thus relying on language-independent information of spoken words representing equivalent
concepts in Dutch and English (Fig. 1A).
The idea of this step was to investigate speech content information in the fMRI responses. To do this, they used a supervised machine learning algorithm combined with single-trial multivoxel classification. Classifications were performed to evaluate whether patterns of voxels conveyed information on the identity of spoken words (within-language word discrimination) and their
language-invariant representations (across-language word generalization).
Within-language word discrimination entailed training a classifier to discriminate between two words of the same language (e.g., horse vs duck) and testing in the same words spoken by a speaker not included in
the training phase. Across-language word generalization was performed by training a classifier to discriminate between two words within one language (e.g., horse vs duck) and testing in the translational equivalent
words within the other language (e.g., paard vs eend), thus relying on language-independent information of spoken words representing equivalent
concepts in Dutch and English (Fig. 1A).