Machine Learning

Question 1

input/output table

Answer 1

represents relationship between inputs (features) and output (target values)

Question 2

what is the goal of machine learning

Answer 2

find underlying function from the data that maps inputs to outputs

Question 3

parameterised hypothesis

Answer 3

model tat is defined by parameters and the learning process is about adjusting these to fit the data

Question 4

hypothesis vs paramterized hypothesis

Answer 4

hypothesis is the proposed function and parameterised is the model that includes parameters adjusted during learning

Question 5

3 types of learning

Answer 5

reinforcement
supervised
unsupervised

Question 6

supervised learning and example

Answer 6

agent learns a function given labelled data inputs and outputs
parameterised model that maps inputs to outputs is chosen and the learning is to find correct parameters that produce correct output
spam/ham

Question 7

unsupervised learning and example

Answer 7

agent is given unlabelled data and has to identify patterns in data
parameterised model is chosen and learning algiorhtm works out parameter values that organises data
used for understanding data
weather patterns

Question 8

reinforcement learning and example

Answer 8

agent learns by interacting with the environment and is given positive/negative feedback based on its actions
learning algorithm wants to chose parameter values that maximise rewards
gets reward and looks back at past actions to see what led to the rewards
AI game playing raj

Question 9

inductive learning

Answer 9

system tries to induce a general rule from a set of observed instances
exam question given in class and put same in exam can’t tell if have just memorised rope learnt or understands

Question 10

hypothesis consistency

Answer 10

consistent if correctly classifies all training examples

Question 11

generalisation

Answer 11

ability of model to adapt to new unseen data not just on training
best models are those that can adapt on unseen data (generalise)

Question 12

overtraining

Answer 12

performs well on training data but poorly on new data

Question 13

overfitting

Answer 13

model becomes too attuned to thee data now which it was trained and loses applicability to any other dataset

Question 14

Occam’s razor

Answer 14

heuristic suggesting choosing simpler ML models as they are expected to generalise better

Question 15

discriminative model

Answer 15

model that attempts to model the agent function directly

Question 16

generative model

Answer 16

model that attempts to model agent function indirectly by modelling the reverse relationship

Question 17

example for reading and writing of discriminative and generative

Answer 17

discriminative can read but not write where as generative has to learn to write before it can read

Question 18

examples of discriminative and generative

Answer 18

generative =bayes classifier
discriminative= decision tree

Question 19

loss function

Answer 19

used to evaluate performance of a model

Question 20

2 types of loss functions

Answer 20

classification error: # mislabeled
MSE: average distance of models output from target value

Question 21

confusion matrix

Answer 21

table used to evaluate performance of classification model
shows TP, FP, TN, FN

Question 22

training vs data sets

Answer 22

split data available for training into 2

training data used to train model/adjust parameters
testing data used to verify performance of the model

Question 23

what does good performance on training and test show?

Answer 23

good on training could be a sign of overfitting
good on performance is a better show of generalisation