Lecture 11

Question 1

machine learning

Answer 1

algorithms that modify themselves through experience and exposure with data. These are statistical equations that adjust their parameters to better classify and predict data

Question 2

artificial intelligence (AI)

Answer 2

algorithms applied to solve novel problems. They utilize algorithm-
refinement from machine learning to understand patterns of data, and then applies a solution in a
novel setting

Question 3

domain specific hypothesis

Answer 3

there are specific brain regions involved for processing specific features

Question 4

spreading activation hypothesis

Answer 4

Activation spreads across interconnected brain regions depending on the features being processed.

Question 5

supervised learning

Answer 5

involves training algorithms by providing datasets with predefined labels. The algorithms attempt to find the best ways to differentiate between these classes of labels.

For example:
what weights can be assigned to highlight the main differences in patterns of activity for natural versus
human made objects?

Question 5

unsupervised learning

Answer 5

involves providing data without class labels. The ML algorithm attempts to
discover the maximum separability between the patterns in the data.

For example: what weights can be
assigned to cluster this data into two different classes with the most separation?

Question 6

testing data

Answer 6

a novel set of data without class labels presented to the ML algorithm
after training. The ML algorithm attempts to “predict” a class label for the data based on the learned patterns of activation from the testing data

Question 6

types of machine learning

Answer 6

(1) support vector machines (SVM)
(2) k-nearest neighbour (kNN)
(3) neural networks (NNs)

Question 7

training data

Answer 7

consists of the data with class labels (or without, if unsupervised). The
ML algorithm devises the optimal weights to differentiate the classes

Question 8

differences between regression and machine learning

Answer 8

1. Machine learning can take advantage of giant swaths of data, taking in large numbers of
   inputs, finding patterns, and providing predictions that maximize accuracy… but so can
   hierarchical regression models
2. Machine learning can optimize its outputs as it gets more experience with data… but so can
   higher level Bayesian statistics

Question 9

support vector machines (SVM)

Answer 9

divides data in multidimensional space into its two classes

Question 10

k-nearest neighbour (kNN)

Answer 10

creates a decision space based on vectors to neighbouring points

Question 11

neural networks (NNs)

Answer 11

a collection of artificial nodes or “neurons” that connect to one another and receive, process, and transmit data, and whose structure is
inspired by biological brains and thus aims to mirror their higher level processing skills

Question 12

hyperplane

Answer 12

divisor in SVM
a multidimensional plane that aims to best divide the data into two class labels

Question 13

support vector

Answer 13

the points that touch the hyperplane, and define the position of the hyperplanes

Question 14

margin

Answer 14

the space between the two hyperplanes that separate the data. The larger the margin, the more
accurate the model’s predictability

Question 15

maximum-margin

Answer 15

the largest possible distance between the hyperplanes separating the data

Question 16

how does k-nearest neighbour (kNN) work

Answer 16

the class identity of a given test point is determined by its closest distance
(i.e., its neighbours) to k number of other points. Training involves calculating vectors between values in
multidimensional space
k is typically a small number (e.g., 3). If k = 1, then the decision of a test point is simply its closest neigbour

Question 16

process of SVM

Answer 16

1. Data is collected (e.g., brain imaging data when people look at
   human made or natural objects)
2. Labelled training data is fed into the SVM algorithm (X × Y x Z x T is
   organized into a single array, labelled as human made or natural)
3. SVM fits a hyperplane to the data through its training. The hyperplane divides the data into two class labels
4. A new set of testing data, unlabeled, is fed into the SVM model
   which attempts to predict the class labels for the test data. Accuracy
   is determined as the percent that was correctly labelled

Question 17

decision space

Answer 17

can be generated by integrating the closest k points in multidimensional space

Question 18

k-fold cross validation

Answer 18

where you combine all iterations with a given subsample of data (k) is left out for testing

Question 19

NNN layers

Answer 19

input layer receives data and begins initial processing.
The output layer outputs the final set of values or set of values

Question 20

simple NNN

Answer 20

one with a maximum of three layers. The input layer is connected directly to the output layer and may have one hidden layer (where additional computations takes place)

Question 21

deep NNN

Answer 21

ave more than one hidden layer. These are often considered black boxes

Question 22

Generative adversarial networks (GANs)

Answer 22

involve two NNs, one that generates an image (the generator), and another that compares it to real images to determine whether it can distinguish between the real and generated one (the discriminator)

Question 23

classifications of neural networks

Answer 23

1. Recognition/classification to learn and identify class labels of information, images, data, etc. This can involve recognizing images such as scenes, people, or objects, or identify things like cancer on a CT scan 2. Generation to produce novel data or images. This is a feature that other ML algorithms like SVM/kNN do not have

Question 24

narrow artificial intelligence (NAI)

Answer 24

generative ai that is trained on data for specific outcomes and can only provide predictions based on its internal boundaries

Question 25

Artificial generalized intelligence (AGI)

Answer 25

can train from a wide variety of contexts and apply its understanding in novel and unseen situations. It can learn from one setting and generalize to a completely different setting. This is on par with the human ability to generalize

Question 26

tokenized

Answer 26

words are tokenized; where they are represented as coordinates in a multidimensional space. A given sentence is thus represented by its relationship between these points in space. ChatGPT then tries to predict the appropriate response based on the structure of these tokens

Question 27

chatgpt

Answer 27

Generative Pre-trained Transformer. These are large language models trained in unsupervised data to generate text. The transformer is what tokenizes words, and these vectors are understood in the context to other tokenized words

Question 28

issues with machine learning

Answer 28

1. Is it picking up actual categorical differences between classes, or some lower-level attribute? 2. Some machine learning (especially DNNs) are black boxes 3. Can normal NHST explain the results? 4. Slight discrepancies in the data can bias the algorithm 5. Overfitting

Question 29

5. Overfitting

Answer 29

A model tries to fit the input data so well that it essentially loses its generalizability to other data

Question 30

4. Slight discrepancies in the data can bias the algorithm

Answer 30

The very fact that the human-made mean is slightly larger, or have a slightly different distribution, may be enough to bias a machine learning algorithm. Often, we have to do lots of transformations to standardize and normalize datasets being compared

Question 31

1. Is it picking up actual categorical differences between classes, or some lower-level attribute?

Answer 31

Differences in relative pixel size, lighting, colour, contrast, illumination, shape, texture, perceived size, perceived weight, emotional valence, colour pallet, familiarity, etc, can all contribute to “differences” between classes. Researchers implementing ML algorithms need to often work harder on standardizing differences between classes than they do on the actual machine learning component. This can greatly restrict and impact the validity of these stimuli

Question 32

2. Some machine learning (especially DNNs) are black boxes

Answer 32

Yes, we can accurately predict information, or even create DNNs that mimic human functions like vision or speech, but do not provide any additional information about what is going on in the brain or with cognition. DNNs can get extremely complex with recurrent networks. It becomes impossible to visualize information at any given layer

Question 33

3. Can normal NHST explain the results?

Answer 33

Fancy applications of fancy algorithms don’t add any explanatory value and can often introduce more problems than solutions. T