Föreläsning 7 (Neural networks 1)

Question 1

What is the artificial analogy of this feature of biological neural nets:
Each neuron sends one electric
pulse

Answer 1

Input neurons

Question 2

What is the artificial analogy of this feature of biological neural nets:
Electrical pulse from the input
neuron is passed on
(propagated) through axon &
axon terminals

Answer 2

Connections

Question 3

What is the artificial analogy of this feature of biological neural nets:
Strength of synapses
determine influence of
input signals on action
potential

Answer 3

Weights

Question 4

What is the artificial analogy of this feature of biological neural nets:
Neuron receiving electrical pulse

Answer 4

Output neurons

Question 5

What is the artificial analogy of this feature of biological neural nets:
Electrical activity in the
soma that if above a
”threshold” outputs an
electric pulse

Answer 5

Activation function

Question 6

What is the simplest type of ANN?

Answer 6

Single-layer perceptron

Question 7

What is the formula to calculate activation value based on the dot product?

Answer 7

𝑥1 ∗ 𝑤1 + 𝑥2 ∗ 𝑤2
x=input value, w=weights
(detta är för exempel 2 inputneuroner, men man plussar alltså summan av input * vikt för alla neuroner)

Question 8

How do you know if you need to update the weights?

Answer 8

1) Compare:
* network output y (ANN prediction), to
* expected output t (ANN target)
2) Update:
* the weights update of the network is
proportional to the network error (t – y)

If the network is working as predicted you don’t need to update the weights!

Question 9

What problems can a SLP solve?

Answer 9

Simpler classification and regression tasks (supervised learning)

Question 10

Vilka två alternativ kan vi välja mellan att träna enkla ANN:s med?

Answer 10

• Manually select the weight values
  ourselves
• Use a learning /training rule

Question 11

Vad är inlärningsregeln för perceptroner?

Answer 11

w’ = w + αx(t-y)