1What is the primary goal of the Anderson et al. (2016) paper discussed in the lecture?
To discover processing stages from EEG data using Hidden Semi-Markov Models and compare them with ACT-R models.
2In the context of Intelligent Tutoring Systems, why is a detailed process model necessary?
To diagnose exactly how a student arrived at an incorrect answer (e.g., identifying a specific procedural error in algebra) rather than just scoring it as wrong.
3What was the conclusion regarding the Psychological Refractory Period (PRP) from the previous lecture?
Behavioral data alone was insufficient to distinguish between serial (ACT-R) and parallel (EPIC) bottleneck models, requiring neural data to resolve the ambiguity.
4What is the main trade-off between EEG and fMRI described in the lecture?
EEG has high temporal resolution (milliseconds) but poor spatial resolution, while fMRI has high spatial resolution but poor temporal resolution (seconds).
5What generates the EEG signal measured at the scalp?
The summed postsynaptic potentials of large populations of aligned pyramidal neurons.
6What is a major limitation of classical Event-Related Potential (ERP) analysis for studying multi-stage cognitive processes?
Trial-to-trial latency variability (jitter) smears the peaks when averaging, obscuring the fine timing details of internal stages.
7In the Associative Recognition task, what is the difference between a Fan 1 and a Fan 2 word?
A Fan 1 word appears in only one studied pair, while a Fan 2 word appears in two different studied pairs.
8Why are reaction times generally slower for Fan 2 targets compared to Fan 1 targets?
Because activation spreads to multiple associations, creating competition and reducing the activation strength for the specific target retrieval.
9In the HSMM-MVPA analysis, what does a "bump" represent conceptually?
The onset of a new cognitive process (or processing stage).
10What is the "linking assumption" connecting the ACT-R architecture to the EEG signal in this specific research?
Each firing of a production rule in the ACT-R model corresponds to a "bump" in the EEG signal.
11How is a "bump" mathematically modeled in the Anderson et al. study?
As a 50 ms half-sine wave.
12What probability distribution is used to model the duration of the "flats" (stages) between bumps?
A Gamma-2 distribution.
13Why is the model called a "Semi-Markov" model rather than just a Hidden Markov Model?
Because the states (stages) have explicit duration distributions, rather than just single-step transition probabilities.
14What dimensionality reduction steps are applied to the raw EEG data before running HSMM-MVPA?
Spatial PCA (keeping ~10 components) and downsampling (e.g., to 100 Hz).
15How many bumps does the ACT-R model predict for "old" targets and "repaired" foils in the associative recognition task?
5 bumps (corresponding to 5 production firings).
16Why might "new" foils have fewer processing stages (and bumps) than targets?
Because the system can quickly detect a lack of memory trace and reject the item without performing a full associative retrieval.
17When comparing the purely bottom-up HSMM to the ACT-R constrained HSMM, which one performed better in cross-validation?
The ACT-R constrained HSMM performed better, despite having fewer free parameters.
18What does the "Leave-One-Out Cross-Validation" (LOOCV) method measure in this study?
How well the model parameters fits the data of a subject that was not used to train the model, preventing overfitting.
19What benefit does "time-warping" trials based on bump locations offer over standard response-locked averaging?
It aligns internal cognitive events across trials, resulting in sharper ERP peaks and revealing components that are otherwise smeared by variability.
20What does the "parietal old/new effect" correlate with in the ACT-R model interpretation?
The activation level of the retrieved memory trace (stronger for Fan 1 than Fan 2).
21In the Sternberg working memory task application, which stage showed the most sensitivity to set size?
The decision/accumulation stage (Stage 4), rather than the memory retrieval stage.
22What effect did word frequency have on the processing stages in the lexical decision task application?
High-frequency words resulted in a shorter duration for the late decision stage (Stage 5), while early visual stages remained constant.
23What are the seven necessary features for the HSMM-MVPA approach to work?
Brain signatures, duration distributions, parameter estimation, bottom-up identification, a process model, a linking assumption, and increased insight.
24What is a limitation of the HSMM-MVPA bump method regarding the total EEG signal?
The bumps account for only about 5% of the variance in the raw EEG signal.
25How does the HSMM-MVPA method help interpret the N2 component in the associative recognition task?
It suggests the N2 reflects response conflict at the onset of the decision stage.