L4

Question 1

ANN

Answer 1

• computing systems inspired by the biological neural networks in human brains
• instead of neurons it has nodes who accept and combine weighted input, add bias, apply a function and output the result
• 3 layers: input, hidden, output
• hidden layer does complex calculations

Question 2

Details of how ANN works

Answer 2

Forward propagation:
- each node has input data, weight, bias and output
- input layer is determined and weights are assigned
- weights help to determine how import each variable is and how much it contributes to output
- inputs are weighted and then summed
- activation function is applied, and if output exceeds given threshold cell is activated and data is passed on to next layer -> feedforward
- output layer predicts final output
Backpropagation:
- in training phase results are fed to the network to compare to own outcomes
- go backward to correct errors and adjust weights

Question 3

Deep Learning

Answer 3

also called deep NN
- for big amount of complex data (normal ML for smaller amounts)
- human-like multilayered processing
- part of ML methods, mix of ANN and representation learning
- handles all types of data
supervised: image classification, object detection, face recognition
unsupervised: word embedding, image encoding
semi-supervised

Question 4

Deep learning procedure

Answer 4

• construct network: input and hidden layers
• train network
• add additional hidden layer on top of previously learned network
• retrain
• repeat until ideal output is generated

Question 5

Representation learning

Answer 5

set of techniques that automatically discover representations needed for feature detection or classification from raw data

Question 6

Application: DL in image processing

Answer 6

• extracts higher level features with each layer -> the more layers the higher the level of features leant
• output layer combines features to make prediction
• lower layers identify edges, high ones things like digits, letters, faces

Question 7

Application: DL in face recognition

Answer 7

with mask:
- feature extraction
- mask detection
- face unmasking
- face restoration
- identification

Question 8

Application: DL in speech recognition

Answer 8

• input audio
• speech signal conversion
• feature extraction
• acoustic models
• pronounciation dictionary
• language models
• recognize words and output in given form

Question 9

other DL applications

Answer 9

Autonomous vehicles
- detecting obstacles and reacting
Recommendation engines
- product, coupons, movies
- stocks and social media