1What is the core distinction between an encoding model and a decoding model in fMRI analysis?
An encoding model predicts brain activity from stimulus features, while a decoding model predicts the stimulus/condition from brain activity.
2What question does a decoding model primarily answer about fMRI data?
It answers whether we can tell what state or input the brain was in from the measured fMRI pattern.
3A model that learns a mapping from brain activity to stimulus labels is known as a(n) _____ model.
decoding
4What does an encoding model specify and learn from data?
It specifies a model of how stimulus features are transformed into neural responses and learns the parameters of this model from data.
5How can an encoding model be used for decoding?
By predicting responses for each candidate input, comparing these to measured responses, and selecting the best match.
6What is a major limitation of a pure decoding model regarding understanding brain function?
It cannot generally be inverted to reveal how information is encoded by the brain.
7In the Kay et al. (2008) study, what biological structures were the Gabor-like filters designed to approximate?
The receptive fields of classic orientation-selective cells in the early visual cortex.
8What was the primary goal of the Kay et al. (2008) encoding model?
To identify which image a subject was seeing from a large, known set of candidate images, based on V1 fMRI activity.
9In the Kay et al. (2008) model, each voxel's response is modeled as a weighted sum of _____-like filters.
Gabor
10What statistical method did Kay et al. (2008) use to learn the weights for each voxel's Gabor filters?
Multiple linear regression.
11How did the Kay et al. (2008) model identify the correct image during the testing phase?
It compared the predicted fMRI pattern for each candidate image with the observed pattern and selected the image with the best match.
12What is the key limitation of the image identification method used by Kay et al. (2008)?
The method cannot freely reconstruct arbitrary images; it can only select the best-matching image from a pre-defined candidate set.
13How did Nishimoto et al. (2011) extend the logic of the Kay et al. (2008) model?
They applied a similar encoding model approach to identify short movie clips from a large candidate set (YouTube clips).
14In the Nishimoto et al. (2011) movie reconstruction, what is meant by an 'implicit prior'?
The reconstructions were biased toward typical movie statistics because the candidate set consisted only of real movies.
15Numerosity-selective neurons, which peak at a preferred number of items, have been found in which two brain regions of macaques?
Frontal and parietal regions.
16Early SVM decoding studies of numerosity showed that classification errors were more common between _____ numbers.
nearby
17What is the term for the topographic mapping of the body surface onto the somatosensory cortex?
Somatotopy.
18The mapping of the visual field across the occipital cortex is known as _____.
retinotopy
19Term: Tonotopy
The topographic mapping of sound frequencies in the auditory cortex, reflecting the organization of the cochlea.
20What is the central hypothesis for why cognitive systems like numerosity might develop topographic maps?
To improve computational efficiency by clustering neurons with similar preferences, thus shortening connection lengths.
21In the numerosity mapping fMRI experiment by Harvey et al. (2013), how was the numerosity stimulus presented over time?
The number of dots on the screen changed slowly and smoothly over time (e.g., from 1 to 20 and back down).
22What task did participants perform during the Harvey et al. (2013) numerosity mapping experiment?
An easy dot-color detection task to ensure they paid attention without explicitly judging numerosity.
23The encoding model for numerosity used by Harvey et al. (2013) was based on what type of mathematical function?
A Gaussian tuning function in numerosity space.
24What are the two key parameters of the Gaussian tuning function used in the numerosity encoding model?
Preferred numerosity (the mean/peak) and tuning width (the standard deviation).
25In fMRI encoding models, what function is used to link the predicted neural activity to the slow blood-flow changes measured by fMRI?
A hemodynamic response function (HRF).
26What was the main finding when the preferred numerosity of voxels was plotted on the cortical surface in the Harvey et al. (2013) study?
They found topographic maps where preferred numerosity changes gradually and smoothly across the cortical surface.
27What was the key conclusion from experiments comparing fMRI responses to dot patterns versus Arabic numerals in numerosity map regions?
These maps encode perceptual quantity (number of items), not the abstract or symbolic concept of number.
28In the study of event timing maps, researchers found patches of cortex that showed tuning to different _____.
event durations
29How do the cortical maps for numerosity and event timing relate to each other?
They are partly overlapping but not identical, suggesting related but partially shared systems for different quantitative dimensions.
30Where are brain regions that respond to symbolic numbers located relative to regions that respond to physical quantities (numerosity)?
They are located in nearby but distinct areas in the parietal cortex.
31Which brain region is specifically activated by quantitative tasks like comparing numbers or solving equations, and is situated between quantity and symbol areas?
A region deep in the intraparietal sulcus (IPS).
32The _____ responds to a wide variety of cognitively demanding tasks, including attention, working memory, and mathematics.
multiple demand network (MDN)
33In the visual short-term memory (VSTM) experiment, what defined the 'load' or task complexity?
The number of unique oriented bars that the participant had to remember.
34What type of responses did voxels show to changing VSTM load, besides simple monotonic increases?
Tuned responses, where the highest response occurred at intermediate loads.
35What did plotting the 'preferred load' across the cortex reveal about the organization of task complexity processing?
It revealed map-like organizations where neurons tuned to similar task loads cluster together in parietal and frontal areas.
36The finding of task-load maps supports the idea that organizational mechanisms from sensory maps have been '_____' for high-level cognition.
recycled
37What is the key 'convolutional' constraint in a DCNN that mimics local connectivity in the visual cortex?
Each unit only integrates information from a small, local neighborhood of inputs from the preceding layer.
38How do the features detected by DCNN layers change with increasing depth?
Successive layers learn to respond to increasingly complex features, from edges to textures to whole objects.
39What surprising representational property emerges in the top layers of DCNNs trained on object classification tasks?
Units often show object-selective tuning (e.g., for faces) that is similar to that of neurons in the inferior temporal cortex.
40What is the technique of modifying an input image to maximally activate a specific DCNN unit or layer called?
Inverting the DCNN (or feature visualization).
41When a DCNN is inverted, what kind of features does the first layer typically add to an image?
Simple oriented edges and lines.
42When a DCNN is inverted, what kind of features do higher layers typically add to an image?
Complex configurations and objects that the network was trained on, such as animal faces or wheels.
43The dream-like images produced by inverting higher DCNN layers resemble reports of _____ hallucinations.
psychedelic
44What theory about human hallucinations is supported by the results of inverting DCNNs?
That hallucinations may involve degraded bottom-up input and amplified top-down feedback from higher-level conceptual representations.
45In the application of 'style transfer', what does a DCNN learn from an artist's paintings?
The characteristic spatial correlation patterns that define the artist's style.
46How does a DCNN for colorization predict plausible colors for a grayscale image?
It uses a high-level representation (object recognition) to infer the likely color distributions based on image context.
47What two types of connections, common in the human brain, are absent in most classical feedforward DCNNs?
Strong within-layer (horizontal) connections and extensive feedback connections from higher to lower areas.
48The encoding model perspective suggests that including known properties of the system, like _____, increases a model's explanatory power.
topographic organization or known tuning functions
49The brain region in monkeys analogous to the human fusiform face area, containing face-selective cells, is the _____ cortex.
inferior temporal
50Research around 2013 began to show that visual area V2 processes patterns in _____ output, not just features in the input image.
V1
51How does the principle of topographic mapping in sensory cortices, like retinotopy, support the possibility of fMRI revealing fine-grained neural properties?
Neurons with similar response properties cluster together, allowing a single fMRI voxel to effectively summarize the local population's tuning.
52What control condition was essential in the Harvey et al. (2013) study to show that tuning was for numerosity and not other visual features?
They repeated sessions with displays that varied non-numerical features like dot size, area, and density.
53What is a primary advantage of using prior knowledge about a system to constrain an encoding model's structure?
It yields more statistically powerful and interpretable models.
54How is the problem of color being unreliably encoded at fMRI resolution handled in the movie reconstructions by Nishimoto et al. (2011)?
Color in the reconstructions reflects the average colors associated with given contrast patterns in the candidate set (e.g., faces appear skin-colored).