AI: Final Exam

Question 1

What are the 3 types of AI capabilities?

Answer 1

Learning, Perception & Cognition

Question 2

What are the types of AI Learnings?

Answer 2

Supervised, Unsupervised & Reinforcement

Question 3

What are the types AI Perception?

Answer 3

Vision, audio, speech and natural language

Question 4

What are the types AI Cognition?

Answer 4

Regression, classification, recommendation, planning, optimization and pattern recognition.

Question 5

AI into business

Answer 5

The push (tech to business) vs pull (business to tech)

Horizontal (similar across industries) vs vertical (unique to industry vertical)

Question 6

What is Machine Learning?

Answer 6

the process of using mathematical models of data to help a computer learn without direct instruction.

Ill - posed problem because a variety of models can explain same data set

Question 7

Overfitting vs Underfitting

Answer 7

Fundamental tension in machine learning due to optimization and generalization

optimization ( adjusting the model to best fit the training data)

Generalization (ensuring the model performs well on data it has never seem before)

Question 8

Overfitting

Answer 8

algorithm has learned “too much” from the training data

Question 9

Underfitting

Answer 9

algorithm has not sufficiently modeled relevant patters in the training data

Question 10

Feature Leakage

Answer 10

features specified in the machine learning model leak information and allow system to “cheat”

Question 11

How do you avoid future leakage?

Answer 11

Model should not have had access to any information about the test set, either directly or indirectly

Question 12

What is the solution to feature leakage?

Answer 12

Split the data into 3 parts “training, validation and test sets”

Question 13

What is another name for test set?

Answer 13

Holdout set. It should never be used to train the model

Question 14

What is deep learning?

Answer 14

incremental, layer by layer approach in which complex representations are developed to map the input to output

Question 15

How do we get an output from an artificial neuron?

Answer 15

2 steps:
1) calculate the weighted sum
2) applying the activation function

Question 16

What are 3 commonly used activations functions in deep learning?

Answer 16

ReLU, Tanh, Sigmoid

Question 17

When do we use ReLU?

Answer 17

ReLu is used for feedforward and convolution neuronal networks

Question 18

When do we use Sigmoid or Tanh?

Answer 18

Used for Recurrent Neural Networks

Question 19

What does a neural network use to update and optimize the weights connecting neurons across different layers?

Answer 19

Error Backprpgation

Question 20

What is a key parameter in error back propagation?

Answer 20

Learning Rate

Question 21

What does a small learning rate lead to?

Answer 21

Leads to insufficient model training

Question 22

Large learning rate

Answer 22

Leads to a quantum mechanic experiment gone horribly wrong

Question 23

When is a pattern linearly classifiable?

Answer 23

1) All data points belonging to one class are on one side of the line
2) All the other data points are on the other side

Question 24

What is more effective in recognizing shifted, shrunk, rotated or deformed images?

Answer 24

CNN (Convolution Neural Network) more effective than feedforward neural network

Question 25

Maxpooling operation

Answer 25

Downsampling operation on each feature map

Question 26

Softmax

Answer 26

Function that generates the probabilities of each case out of multiple cases

Question 27

What are CNNs capable of?

Answer 27

Learning local, translation invariant features; highly data efficient on perceptual problems

Question 27

What are CNNs capable of?

Answer 27

Learning local, translation invariant features; highly data efficient on perceptual problems

Question 28

Area Under Curve (AUC)

Answer 28

Health applications - an important measure of the accuracy of of a binary classifier

Question 29

Classifier types

Answer 29

Perfect has an AUC of 100% , random classifier has AUC of 50% **Higher AUC indicates a better classifier

Question 30

What is MLOps?

Answer 30

Software development approach that brings AI development into business operations together

Question 31

What are disadvantages of the Waterfall Model?

Answer 31

Specifies all the requirements at the begining of the project and allows little change during the development process

Question 32

What is the agile approach?

Answer 32

Designed for changes during the dev process

Question 33

What is Scrum?

Answer 33

a type of agile approach that organizes a project into a series of sprints

Question 34

What is the central problem of deep learning?

Answer 34

Overfitting which occurs when the accuracy on the training exceeds that on the data set

Question 35

What are some key strategies to fight overfitting?

Answer 35

Get more training data, reduce the size of the neural network (pruning), add weight regularization, add dropout

Question 36

Sequence Learning

Answer 36

one to one, too manny, or many to many

Question 37

What do simple RNN model include?

Answer 37

Short-term memory input and output for each neuron (does not reflect long-term tendencies)

Question 38

Long Short Term Memory (LSTM) RNNS

Answer 38

Include carry units to ensure long term tendencies that are not readily erased during the learning process

Question 39

How do you give RNNS the ability to keep track of long-term dependenices?

Answer 39

Long Short Term Memory

Question 40

What is a reinforcement learning model?

Answer 40

Learns by doing, agent who takes a sequence of actions within some environment and receives direct feedback

Question 41

What is a Q learning model?

Answer 41

Define the quality of an action by how rewarding it is. 1) instant reward 2) indication of future reward

Question 42

What are the key parameters in a Q learning model

Answer 42

Learning rate exploration rate discount rate

Question 43

Q learning model - Learning rate

Answer 43

Determines to which extent newly acquired info overrides old information

Question 44

Q learning model - Exploration rate

Answer 44

Determines the probability of taking a random action rather than an action that gives maximum value

Question 45

Q learning model - Discount rate

Answer 45

Determines the importance of future reward vs instant rewards

Question 46

Neglect tolerance

Answer 46

The amount of time that a robot can be neglected

Question 47

Interaction time

Answer 47

The avg. amount of time it takes for human operators to interact with the robot (or AI system)

Question 48

Level of Autonomy

Answer 48

A system that can be neglected for a long period of time without interaction

Question 49

Fan out

Answer 49

The upper bound on the number of robots that a single person can manage