ml

Question 1

How do neural networks work?

Answer 1

Input layer

1+ hidden layers

Output layer

all neurons in each layer connected to all neurons in the next layer “fully connected”

Each neuron adds up the inputs, puts the sum through an activation function

Each connection has a weight which multiplies the value by a certain amount

Question 2

What are the 5 categories of neuron in CNNS?

Answer 2

Input - usually in 2d grid for image input

CONV -Layer that performs a convolution

RELU - activation function

POOL - sample down or average the inputs

Fully connected - usually at output stage

Question 3

How does K-means work?

Answer 3

Assign feature vector to class based on the current means of each class (feature vector takes the closest class as it’s class)

Update the location of each class mean after all feature vectors have been assigned

Question 4

Benefits/drawbacks k-means

Answer 4

Global optimum not guaranteed

Stochastic

but very fast to run

Question 5

What is wisard?

Answer 5

Basic pattern recognition

Uses blocks of memory

Binary images

Choose random pixel patterns and set to 1 and 0 based on the pixel values

Testing: take the input and test over many different locations, gives a score

Highest score of all trained recognisers wins :P

Question 6

What are SVMs?

Answer 6

Linear classifiers

Can be made to do non-linear using kernel tricks

finds the hyperplane that best separates the two classes

Deterministic