Exercise 5 - Hebbian Learning

Question 1

3 generations of neuron models

Answer 1

1st generation: binary
2nd generation: real numbers
3rd generation: action potential

Question 2

learning

Answer 2

finding the right weights to solve the task

Question 3

universal approximation theorem

Answer 3

• a neural network with only a single hidden layer can approximate any function arbitrarily well

Question 4

what learning model is hebbian learning?

Answer 4

unsupervised learning

Question 5

Two learning rules in hebbian learning

Answer 5

rate-based hebbian rules

precise-timing-based rules

Question 6

Hebbian learning

Answer 6

When an axon of cell A is near enough to excite B and repeatedly or presistently takes part in firing it, some growth process takes place such that A’s efficiency is increased.

Question 7

Properties of hebbian learning

Answer 7

• saturation: avoid unbound growth of synaptic weight
• competition -> selectivity: avoid weights to converge to the same value
• locality: weight change depends on local variables

Question 8

Does Oja’s rule satisfy all properties?

Answer 8

Yes

Question 9

Does covariance learning with sliding threshold fulfil all properties?

Answer 9

Yes

Question 10

Spike-timing dependent plasticity (STDP)

Answer 10

• temporally asymmetric form of hebbian learning induced by tight temporal correlations between the spikes of pre-and postsynaptic neurons