Which of the following statements are false with reference to self$-$organizing feature maps? Incorrect

answers will be penalized. Write down only the letter$($s$)$ of your answer$($s$)$

$($a$)$ Training minimizes the mean$-$squared error

$($b$)$ Self$-$organizing feature maps use a competitive training approach

$($c$)$ They are robust to skew class distributions

$($d$)$ The best approach to weight initialization is to initialize the weights by sampling weight values from

a uniform distribution in the range $-0\mathrm{.}1,0\mathrm{.}1$ to ensure that initial weights have small values

$($e$)$ Training is a computationally expensive process

$($f$)$ Batch training will reduce the computational cost

$($g$)$ Can be used for regression problems

Explain how a self$-$organizing map $($SOM$)$ can be used to profile customers of a medical aid company. $(5)$

True or false: after training a selforganizing feature map, output neurons that win for similar inputs are

usually far apart from each other in the map?

Training of a SOM is computationally expensive. Discuss three approaches to reduce the computational

cost of training a SOM.

Can a SOM be used as a classifier? Motivate your answer.

Explain how a SOM can be used for data imputation.

Unsupervised learning algorithms do not use target output values. So$,$ what does unsupervised learning

learn?

You are using a self$-$organizing feature map to develop a recommender system for Netflix. Only descriptive

features characterising movies are used. Assume that the self$-$organizing feature map has been trained on

a large number of movies.

$($a$)$ After watching one movie, how can the trained map be used to recommend the next movie?

$($b$)$ After having watched a large number of movies, how can the trained map be used to build a profile

of the types of movies watched?

$($c$)$ In unsupervised learning, there are not target output values. So$,$ what does unsupervised learning

learn?

Consider a learning vector quantization I $($LVQ$-$I$)$ algorithm used for data clustering. Propose

an approsch to determine the optimal number of output units, i$.$e$.$ the optimal number of clusters, during

training of the LVQ$-$I.