1What are typical applications of machine learning mentioned in the text?
Image labelling, game opponents, weather prediction, autonomous driving, and recommendation engines.
2According to Samuel (1959), what is the definition of machine learning?
The field of science that gives computers the ability to learn without being explicitly programmed.
3What are the three components of Mitchell's (1998) definition of machine learning?
Experience (E), Task (T), and Performance measure (P).
4What is the main conceptual advantage of Mitchell's definition of machine learning over Samuel's?
It makes no cognitive assumptions and expresses learning in operational/measurable terms.
5What defines Supervised Learning?
The training data consists of pairs of input and desired output (labelled examples), and the goal is to map new inputs to correct outputs.
6What is the difference between Classification and Regression in supervised learning?
Classification predicts discrete outputs (e.g., spam vs. not spam), while Regression predicts continuous outputs (e.g., amount of rain).
7Why is simple accuracy a problematic metric for classification when classes are imbalanced?
Because a trivial classifier (e.g., "always spam") can achieve high accuracy without actually learning to distinguish classes.
8What is Leave-One-Out (LOO) cross-validation?
Training on N-1 examples and testing on the remaining 1, repeated for all N examples.
9What defines Unsupervised Learning?
There are no desired outputs or labels; the goal is to find structure in the inputs (e.g., clustering).
10How does Reinforcement Learning differ from Supervised Learning?
Feedback is delayed and coarse (e.g., "did I win?"), rather than dense step-by-step labels.
11When should one use a generic machine learning classifier versus a specific statistical model?
Generic classifiers are for when interpretation is low priority; specific models are better when testing specific hypotheses or when interpretability is central.
12What is the key contrast between ML classification and Classical Statistics regarding output?
ML tells you *that* a relationship exists (high-dimensional), while Statistics tells you *which* variables matter and *how* they affect the output.
13What is the limitation of traditional univariate fMRI analysis ("blob" analysis)?
It shows where activation changes (magnitude) but misses information contained in the pattern of activity across voxels.
14What is the difference between the "Average" approach and the "Pattern" approach in fMRI analysis?
The Average approach looks for changes in mean signal (univariate), while the Pattern approach looks for changes in the distribution/arrangement of signals even if the mean stays the same (multivariate).
15How does a Support Vector Machine (SVM) classify data?
It finds a decision boundary (hyperplane) in high-dimensional space that separates data points into classes.
16What is a major weakness of MVPA (Multi-Voxel Pattern Analysis) described as the "Bag of Voxels" problem?
It treats voxels as undifferentiated dimensions, ignoring the spatial structure and anatomy of the brain.
17Why was decoding orientation from V1 voxels considered a paradox?
Because standard fMRI voxels (3mm) are much larger than the fine-scale orientation columns they contain, so the signal was expected to average out.
18What are the two explanations for why coarse fMRI can decode fine-scale orientation?
1. Random sampling/aliasing (idiosyncratic biases in voxels). 2. Large-scale biases (maps of preferred orientation across the cortex).
19In the context of the vegetative state study, what did MVPA reveal?
That some patients could modulate brain activity to match instructions (tennis vs. navigation), providing evidence of awareness and communication capacity.
20What is Cross-Classification?
Training a classifier on data from one context (e.g., awake state) and testing it on data from another (e.g., sleep state) to test for representational stability.
21What is the core idea of Representational Similarity Analysis (RSA)?
Focusing on the *distances* (dissimilarities) between conditions rather than the boundaries between classes.
22What is a Representational Dissimilarity Matrix (RDM)?
A matrix where rows and columns represent stimuli, and each cell contains the dissimilarity score (distance) between a pair of stimuli.
23What is the major advantage of RSA when comparing different data types (e.g., human fMRI vs. monkey recordings)?
It abstracts data into a common "language" of dissimilarities, allowing direct comparison of representational structures across different units and species.
24What is Multidimensional Scaling (MDS) used for in the context of RSA?
It projects high-dimensional dissimilarity data into a low-dimensional space (usually 2D) to visualize clusters and relationships between stimuli.
25When is Machine Learning described as "lazy statistics"?
When it is used as a flexible tool to detect that a relationship exists without identifying exactly what that relationship is.