College #3 - Neuroimaging of the semantic brain network Flashcards
Perisylvian language network
The perisylvian language network is responsible for the production and comprehension of language. Is found due to impairment of linguistic abilities after brain damage, activation in fMRI research and ERP evidence thanks to EEG.
Functional Magnetic Resonance Imaging (fMRI)
Measures the blood oxygenated level dependent signals (BOLD). So it looks at oxygenated vs non oxygenated blood.
Voxel
A 3D pixel of a ‘photograph’ of the brain. In each voxel the blood oxygen level is measured.
General linear model
Voxelwise multiple regression model.
procedure:
- Create design matrix
- Fit model
- Statistical inference
This model is at subject level, but can also been done at group level.
Group level interference
Does condition 1 significantly activate the brain?
B1 > 0
Is the difference between condition 1 and 2 significant?
B1 - B2 > 0
B2 - B1 > 0
Limitations of Voxelwise approach
- Voxels are modeled independently; gives a lot of data and you can’t see the interaction.
- Amplitude changes can arise for many different reasons.
- Mass univariate testing.
- Depends heavily on linear model.
> > This leaded to focus on pattern analyses; groups of voxels.
Multivariate pattern analyses
Study of distribution of activation across multiple voxels.
Main assumption: Cognitive processes are not limited to isolated voxels, but occur in larger units.
Pattern definition: Area under curve of the trial-specific BOLD-response or the beta-estimate.
Interpretation and analysis:
- Dimensionality reduction
- Classification analysis
- Representational similarity analysis
Classification analysis
= Decoding neural patterns.
A machine learning algorithm assigns pattern to predefined stimulus identity or category. The goal is classification and cross-validation.
Data gets split in trainingsdata with labels and test-data. Out of the trainingsdata a trained classifier gets formed. The clarifier determines labels for unseen data. The accuracy of the classification of the test-data determines whether there is something in the neural patterns that allow for classification.
Represented similarity analysis
Is used to analyse neural patterns. It is the correlation between two similarity matrices (behavioral and neural).
It is a second order similarity test; it computes the similarity of similarities.
First-order similarities - neural
Procedure:
- Identify number of regions.
- Make matrices.
- Vectorisation.
- Look for similarity between the regions/vectors.
First-order similarities - behavior
> > Semantic models
- The word meaning is represented as Vector.
- Semantic similarity = cosine similarity between word vectors.
- It is association-based, and distributional.
> > Other models
- Concreteness_a - Concreteness_b
- Visual
Permuation testing
Is looking for statistical interference.
This is done by repeatedly randomizing 1 of the matrices and calculating the correlation. In this way you generate a distribution of randomized correlations. Check whether your result is significant»_space; Does it fall within the top 5% of this new distribution.
Adavantages of representation similarity analysis
> > Relates info from different spaces
- Connects brain and behavior.
- Allows for between subjects/species analysis
- You can evaluate computational models of the brain.
> > No A priori assumptions
Confirmation of hypotheses about where info is represented in the brain.
Fine-grained; Works at stimulus level.
What is a network?
Consists of nodes and edges.
Nodes (vertex) = Fundamental units / main building blocks.
Edges = Connections between the nodes. These edges determine the flow of information. They constrain how the information transfer between nodes can happen.
- Directed or undirected (1 way or 2 way).
- Binary or weighted
Two connected nodes are called adjacent nodes.
Characteristics of brain networks
Regular brain network = specific.
Small-world brain network = Balanced/broad abilities.
Random brain network = General.
FOTO INVOEGEN
