Machine learning

Question 1

What is AI?

Answer 1

AI = generic term for machine learning

Question 2

Going from generic to more specialized fields, which of the following is true:

a) AI is a form of machine learning, which is a form of deep learning
b) Deep learning is a type of AI, which is a type of machine learning
c) Machine learning is a type of AI, which is a type of deep learning
d) Deep learning is a form of machine learning, which is a form of AI

Answer 2

d) AI → ML → deep learning

Question 3

What kind of learning is artificial intelligence?

Answer 3

Rule-based learning

Can think of in terms of if/else or if/then statements

Question 4

On what is machine learning based?

Answer 4

Machine learning is based on examples

Comprises training & prediction stage

Question 5

Forms of learning frameworks

Answer 5

1) supervised (learning with a labeled training set)
2) Unsupervised (finding patterns in unlabeled data)
3) Reinforcement learning (learning based on feedback)

Question 6

What do labels do?

Answer 6

Help to define the research question

Question 7

In supervised learning, the machine…

Answer 7

learns functions that maps inputs to outputs

training is in the form of example inputs and desired outputs

Question 8

In unsupervised learning, the machine…

Answer 8

learns structures within the data
-→ the goal is to discover hidden patterns and be a means toward and end
(feature learning)

ex: classifying different types of cats w/o specific characteristics of each cat type but rather the general classification of cats ; ML can then detect variability w/in the data classes

Question 9

Which of the following is/are true regarding feature-engineering in deep learning and ML?

a. In deep learning there is no prior feature extraction, meaning the end result is the own representation of data
b. both standard ML and deep learning have features that are hand-crafted (ex: the distribution of white & gray matter, volumes, etc.)
c. only deep learning uses hand-crafted features
d. only standard machine learning uses hand-crafted features
e. both standard ML and deep learning have no prior feature extraction

Answer 9

a, d

Standard machine learning involves hand crafted features and deep learning has no prior feature extraction

Question 10

What do generative models model and what are some examples?

Answer 10

model p(x,y) with x features and y labels

examples: Native bayes, gaussian mixture models, generative adversarial networks, etc.

Question 11

What do discriminative models do and what are some examples?

Answer 11

Model p(y | x) w/ x features and y labels

examples: decision trees, logistic regression, neuronal networks, Gaussian processes, vector machines, etc.

Question 12

On what is deep learning based? and on what does it rely?

Answer 12

Based on end-to-end learning

Relies on basis of artificial neural networks

Does not need prior feature extraction to come to a decision

Question 13

What are the different architectures of artificial neural networks (ANN)?

Answer 13

• Fully connected neural networks (i.e. MLP)
• Convolutional neural networks
• Recurrent neural networks
• Generative adversarial networks

Question 14

MLP structure

Answer 14

comprises input layer, hidden layer(s), and output layer

The input layer received inputs of different weights (w) based on node connectivity

Hidden layer performs non-linear transformation of the inputs and provides input to the output layer

Question 15

What kind of learning does the MLP use? How?

Answer 15

Uses supervised learning in the form of backpropagation

The weight of the inputs is changed so that the error decreases = training

Learning = optimization

Question 16

What are convolutional neural networks based on?

Answer 16

Are deep learning networks that are optimized for processing arrays (often in the form of videos or images)

Question 17

Convolutional neural network layers

Answer 17

Convoluting
Pooling
Non-linearity

Question 18

Neural networks are what kind of model?

Answer 18

Discriminative

Question 19

How to overcome the issue of small sample size

Answer 19

Transfer learning (a form of data augmentation)

i.e. train models with other, previous models

Question 20

How to overcome the explainability/black box issue

Answer 20

Can overcome the black box issue (idea that method to receive results is secret or difficult to understand) by generating heat maps based on input data

Step 1: train model & apply to test data
Step 2: generate heat map for each subject in the test set

Question 21

Challenges w/ machine learning & AI

Answer 21

sample size, explainability, bias & confounding variables not accounted for, translation into clinical practice, reproducibility and benchmarking