AIML

Question 1

What is supervised learning, and why is it used?

Answer 1

Learning with labels attached. Trained to generalize new features in the dataset.

Question 2

What is unsupervised learning, and why is it used?

Answer 2

Learning with features but no labels. Trained to predict future information, using a hypothesis of equation y = p1(x) + p0.

Question 3

What is cost?

Answer 3

Poorness of fitted line to data.

Question 4

What is the K-NN algorithm?

Answer 4

Supervised learning algorithm to classify new incoming data with low complexity

Question 5

What are three advantages of the K-NN algorithm?

Answer 5

Simple to implement.
Flexible to all features and distance equations.
Easily handles multi-class data.

Question 6

What are three disadvantages of the K-NN algorithm?

Answer 6

Large search problem to find nearest neighbors, can be intensive
Requires a large amount of stored data with many classes
Distance function should be meaningful

Question 7

How does K-NN operate?

Answer 7

Choose some K as the number of neighbors to take.
Locate K nearest neighbors for unclassified example. Should not be multiple of K. Optimize K on observation.
Most votes wins.

Question 8

What is the K-means Clustering algorithm?

Answer 8

Unsupervised learner for classification or regression
Finds K groups in the set, defined by centroids.
Guaranteed to converge on a result, though it may not be local optimum

Question 9

What are three advantages of the K-means clustering algorithm?

Answer 9

Can be used for any kind of grouping
Thanks to the simple layout of data, new data can easily be applied to a cluster
Clustering allows finding groups that have formed organically without definition

Question 10

What are three disadvantages of the K-means clustering algorithm?

Answer 10

Cannot handle outliers
Cannot handle complicated cluster types
Cluster assignments change with each run

Question 11

How does K-means clustering work?

Answer 11

Randomly generates K centroid locations in result space
Each point assigned to nearest centroid
Centroid relocated to mean of all assigned points
Iterates until stopping, due to no change or sum of distances minimized, iteration cap etc.

Question 12

What is the search space?

Answer 12

Graph representing how good each solution is, attempting to find the global optimum (vs. the local optimum)

Question 13

What are the components of a neuron?

Answer 13

Synapses, receiving numerical input
Summation sub-unit sums weighted inputs to single value
Activation sub-unit maps to new output
Output through axon to all connected neurons

Question 14

What is a hidden layer?

Answer 14

Layers between the input and output.

Question 15

What is feed-forward?

Answer 15

The input is fed through all connections, weighted and passed forward. An output is produced.

Question 16

What is linear seperability?

Answer 16

Whether a single straight line can be drawn that would separate all positive and negative values. XOR is not linearly separable.

Question 17

Define bias vs variance.

Answer 17

Bias is a simple model which fails to capture relationships within the data.
Variance is a complex model that has overfitted to the data, producing noisy results.

Question 18

What is regularisation?

Answer 18

Group of techniques coercing a model to infer simpler results, aiding the bias-variance trade-off.

Question 19

Give three examples of regularization with brief descriptions.

Answer 19

L1 - Prevents overfitting by applying penalty to loss, driving unimportant values to zero. Smaller models.
L2 - Adds extra term to loss. Drives towards smaller values.
Dropout - Randomly shuts off neurons. Forces robust learning.

Question 20

What is a genetic algorithm?

Answer 20

Probabilistic search method adopting natural selection approach

Question 21

List three benefits of a genetic algorithm.

Answer 21

Supports multi-objective optimization
Good for noisy data
Always provides an answer
Easily parallelable

Question 22

How does a genetic algorithm represent solutions?

Answer 22

Encoded, often using binary bit strings (chromosomes)

Question 23

Describe the broad steps of a genetic algorithm.

Answer 23

Initialize a population of solutions
Evaluate each using an objective function
Create new solutions via selection, crossover, mutation and elitism
Replace old population and repeat

Question 24

Define selection in a genetic algorithm.

Answer 24

Each solution arranged by fitness value
Total value calculated
Each solution calculates its % of total fitness
Biased roulette wheel built from probabilities

Question 25

Define crossover in a genetic algorithm.

Answer 25

From selection roulette wheel, choose two parents and switch one or more pairs of bits

Question 26

Define elitism in a genetic algorithm.

Answer 26

Select some best candidate solutions and grant free passage without modification

Question 27

Define mutation in a genetic algorithm.

Answer 27

Random alteration or flip of a bit in a solution.