Fundamentals of MRI
Water moves preferentially in one direction due to barriers (e.g., bundled axons). indicates the presence and orientation of fiber tracts (showing white matter)
Water moves equally in all directions (e.g., ventricles; gray matter)
What is anisotropic diffusion?
What is isotropic diffusion?
Visualize bundles of axons that connect distinct regions
*i like to think about it as a highway where it traces the paths of white matter fiber tracts (bundles of nerve fibers) by following the direction in which water molecules move
Tractography
Instead of testing individual [], MVPA analyzes distributed activation across multiple of these.
What is voxels?
This foundational learning principle suggests that neurons that fire together wire together. In MVPA studies, it explains why repeated co-activation of the same neural pattern leads to stronger within-item similarity across repetitions — predicting better memory.
What theory does this concept support?
What is Hebbian learning?
What is the simple strengthening theory?
The similarity or dissimilarity of multivoxel patterns across conditions or stimuli.
What is representational similarity analysis (RSA)?
This scalar measure, commonly used in diffusion MRI, reflects how strongly water diffusion is directionally constrained within white matter tracts.
A drop in this [] may indicate damage to these neural structures.
What is fractional anisotropy (FA)?
What are axons or myelin?
the starting point in tractography, placed where a known white matter fiber pathway is likely located.
Often combined with an "AND" constraint: only show fibers that also pass through a second region
What is a seed voxel and how is it used in tractography?
In this MVPA approach, researchers test whether patterns of voxel activity from one part of the data (like even-numbered runs) resemble those from another (like odd-numbered runs).
This MVPA approach uses machine learning to classify brain activity patterns based on labeled conditions.
Provide examples for both approaches.
What is the correlation approach? For example, seeing if "chair" patterns in one half of the data match "chair" patterns in the other half.
What is the classification approach? For example, training a model to distinguish between patterns evoked by viewing shoes versus chairs, then testing it on new trials.
This model predicts brain activity based on a given stimulus or behavior. An example of this is neural prosthetics such as mapping brain activity to control an artificial limb.
However, this model reverses the logic and predicts the stimulus or mental state from brain activity patterns. An example of this in an EEG lens is ERP-based spelling device such as decoding intended letters from brain responses to flashing options.
What is encoding model?
What is decoding model?
In Kriegskorte & Kievet, 2013, the researchers used RSA with a subject viewing faces or houses. While the primary and secondary cortex exhibit differences in neuron responses, what can it not tell us?
What is clearly not distinguishing between faces and houses as it just shows us the cortical areas that show distinct activity patterns when processing faces versus houses.
This form of connectivity is based on the physical wiring of the brain, such as white matter tracts. uses diffusion imaging
This form of connectivity is inferred from correlated BOLD signal fluctuations between brain regions.
What is structural connectivity?
What is functional connectivity?
These two neural tracing techniques help map brain pathways:
-one uses a tracer taken up by cell bodies and transported along axons to show outputs.
-the other uses a tracer taken up at axon terminals and transported back to cell bodies to show inputs.
What are anterograde tracing and retrograde tracing?
This study uses MVPA in a study to answer what we already know about vision.
Even though voxels contain hundreds of thousands of neurons, MVPA can decode stimulus features like edge orientation because of this subtle property of cortical column distribution. ANSWER THIS QUESTION
This suggests that while the map of orientation-sensitive voxels differs between subjects, classification remains stable [] a subject, but not [] subjects.
What are small irregularities in the spatial layout of orientation columns?
What is within a subject, and not between subjects.
An fMRI study trains a decoding model to identify which emotion a participant is imagining based on voxel activity patterns. The model performs well, using data from over 30,000 voxels — but researchers warn that its success might reflect statistical overfitting rather than true emotional representation. This concern stems from having more features than trials and relates to this fundamental problem in high-dimensional MVPA analyses.
What is the solution to this concern?
What is the issue of dimensionality?
What is splitting the data into separate chunks (like different runs) and looking for consistent condition differences across these chunks?
In the Saarimäki et al. (2016) paper, the researchers used the classification approach in two formats to test the data.
Then, researchers used RSA to compare how similar emotional states felt to participants and how similarly they were represented in the brain. What key result was required for RSA to be considered successful?
What is within-subject classification (training of half of a subject's trials and test on the other half) and a between-subject classification (train on all but one subject and testing on the missing subject)?
What is a significant correlation between the neural similarity matrix and the behavioral similarity matrix?
Though fractional anisotropy (FA) is widely used to infer white matter integrity, studies warn that its interpretation is complicated by this factor.
What is anatomical variability?
Researchers rely on [] in creating the “AND” gates to include only plausible paths and “NOT” gates to exclude unlikely ones. What key strategy do these techniques rely on?
What limitation would occur if research did not use these "AND" and "NOT" gate methods?
What is using prior anatomical knowledge to constrain tracking?
To reduce false positives in tractography
Success in MVPA classification does not imply that the brain itself uses the same representation; to support this claim, researchers must ensure it reflects the experimental conditions, and not idiosyncratic association, by testing this accuracy across subjects, sessions, or similar conditions.
What is generalization?
What is the central question all three papers are testing with MVPA?
This hypothesis suggests that consistent neural activation patterns across repeated study sessions enhance memory retention. This paper is highlighting this theory.
According to this hypothesis, variability in neural activation during repeated retrievals can lead to improved memory performance. This paper is highlighting this theory.
What do these conflicting findings from the two papers suggest about memory encoding processes?
What is how does repetition improve memory retention?
What is the strengthening hypothesis (Xu et al., 2015)?
What is greater dissimilarity (or variability) in neural patterns (Wirebring et al. 2015)?
What is may depend on the brain region involved and the nature of the task? Retrieval tasks involving parietal regions might benefit from variability, implying that different types of information or cognitive tasks require distinct encoding strategies
Representational Similarity Analysis (RSA) was utilized in the Xu paper across the three experiments. How was RSA applied in the studies, and what were the findings?
Exp 1 and 2 (had recall for second): Pattern similarity (significance in the right inferior parietal louble and right lateral occipital complex) in showing more similarity across repetitions for remembered items compared to forgotten items
Exp 3: left LOC showed stronger within-item similarity for remembered words, and provides more targeted evidence that repetition can lead to stronger memory traces, particularly when similarity is assessed appropriately, and task design supports spacing effects
These ultra-slow (< 0.1 Hz) fluctuations in the BOLD signal, observed even during rest, serve as the foundation for identifying functional brain networks without any task.
What form of network do these refer to?
What are spontaneous low-frequency fluctuations, as first shown by Biswal et al. (1995)?
Default Mode Network
Researchers use tractography to estimate white matter connections, but why are longer-range tracts particularly prone to error, and what does this reveal about the nature of tractography as a method?
What is that longer tracts accumulate more uncertainty and signal loss, revealing that tractography is an inference-based method rather than a direct measure of structural connectivity?
There are a few issues with MVPA.
One is that MVPA classification might appear successful, but this concern warns that the brain pattern could reflect unintended thoughts or associations — like a trial reminding the participant of their cousin Cindy.
Also, MVPA can identify differences in brain activity between two cognitive tasks, but it does necessarily discriminate between the [] that explain those tasks.
What is the problem of interpretability in MVPA?
What is the underlying theories?
In a common modeling framework of brain information flow by Kriegeskorte & Douglas, 2019, the researchers believe our behaviors are achieved based on this timeline. What is the correct order of processing in this model?
What is: World → Area A → Area B → Area C?
Area A encodes sensory input, Area B transforms or recodes that information, and Area C generates an output like a decision or action.
In studies using RSA and classification, cortical regions often yield higher decoding accuracy than subcortical ones. What is one reason this difference may not reflect better representation, but rather a methodological advantage?
What is that cortical regions often contain more voxels or have more reliable fMRI signal, which can inflate classifier accuracy?