12. What is Hebbian learning which underlies the core idea of weight changes?
Characterised by 'neurons that fire together wire together'
Finding a set of connection weights to make the network useful by making systematic changes to the weights in response to input patterns
A concept analogous to the total strength of all inhibitory and all excitatory connections between the neurons
Neural networks can be connected to transform ANY set of input patterns into ANY set of output patterns
13. What variant of learning does deep learning use?
Supervised learning
Unsupervised learning
Competitive learning
14. Do neural networks start with initially random weights?
Yes
No
15. Which of these characterises unsupervised learning?
Network learns to identify clusters/statistical structures in the input. Output neurons compete to respond to each input pattern (lateral inhibition)
Network learns by discovering the inherent structure in the input. Based heavily on hebbian learning
Network learns by being presented with example input patterns and the desired output pattern. Initial output is different but is gradually and systematically changed to reduce discrepancy.
16. In supervised learning, what happens after each input pattern?
Network checks the output against the underlying structure of the input to attempt to reduce discrepancy in the whole model
Connections are strengthened as to reduce contributions to discrepancy between observed and desired output
Connections are weakened/strengthened as to reduce contributions to discrepancy between observed and desired output
Connections are weakened as to reduce contributions to discrepancy between observed and desired output
17. What is the concept of generality?
Neural networks can be connected to transform ANY set of input patterns into ANY set of output patterns
A concept analogous to the total strength of all inhibitory and all excitatory connections between the neurons
Finding a set of connection weights to make the network useful by making systematic changes to the weights in response to input patterns
Characterised by 'neurons that fire together wire together'
18. What are the two criteria for generality?
Each input pattern maps onto ANY output pattern and there are enough units (4) and layers of units (unlimited)
Each input pattern maps onto ANY output pattern and there are enough units (unlimited) and layers of units (4)
Each input pattern maps onto only ONE output pattern and there are enough units (unlimited) and layers of units (4)
Each input pattern maps onto only ONE output pattern and there are enough units (4) and layers of units (unlimited)
19. Which is correct?
(sum of)Activation of input neuron x output
(sum of)Activation of input neuron -output
(sum of)Activation of input neuron x weight of input
(sum of)Activation of input neuron - weight of input
20. Which of these characterises supervised learning?
Network learns to identify clusters/statistical structures in the input. Output neurons compete to respond to each input pattern (lateral inhibition)
Network learns by being presented with example input patterns and the desired output pattern. Initial output is different but is gradually and systematically changed to reduce discrepancy.
Network learns by discovering the inherent structure in the input. Based heavily on hebbian learning
