Computational Neuroscience 2

Question 1

Learning is achieved by…

Answer 1

…modifying the synaptic weights of the network

Question 2

First learning algorithm that automatically changed the weights in neural networks came from…

Answer 2

…Widrow-Hoff trying to devise an automatic speech to text system

Question 3

Widrow-Hoff rule

Answer 3

1. Initial weights can have random, or zero, values
2. A CS is presented and the output recorded
3. The output is compared to a desired output, and if there is an error the weights are adjusted in order to reduce the error (supervised learning).
   Same as the Rescorla-Wagner models.

Question 4

Extinction trials

Answer 4

Trials in which the learned condition is ‘unlearned’. That is achieved by exposing the animal to the CS without the US.

Question 5

The blocking effect

Answer 5

Pavlov suggested that it was only the co-occurrence of the CS and US that were key to conditioning.
However, by the late 60s several studies showed that in order to produce conditioning, an stimulus needs to have predictive power about the US.
Rabbits are separated in two groups: pre-trained group and sit-exposure group
Phase 1
• The pre-trained group is trained to the tone as CS before the airpuff US
• Meanwhile the sit-exposure animals sit in the experimental chamber.
Phase 2
• Both groups are trained to respond to a CS consisting of simultaneous tone and light (read the paper)

Question 6

Error correction learning in the brain

Answer 6

The Rescorla-Wagner model requires a teaching signal to act as a feedback mechanism.
Therefore, it is not a model of learning at neuron level
However there is evidence that certain brain structure do provide a feedback signal that the system can used as a teaching signal.
• One example is the cerebellum — an area heavily involved in learning reflex responses
• In 1968, Richard Thompson suggested that there was a teaching signal in the cerebellum that the brain could use in conditioning

Question 7

Error correction in the cerebellum

Answer 7

CS information reaches the cerebellum
Information about the US (airpuff)
travels through the inferior olive
Climbing fibers carry this information to the cerebellum
The output of the cerebellum is related to the conditioned response (eyeblink)
Part of the output also travels back to the inferior olive through inhibitory connections.
The pathway activated by the US acts as the desired output.
Thus the difference between the CR and the US determines the activity of the inferior olive
Inferior Olive activity = US - CR

Question 8

Generalisation

Answer 8

The capacity to learn stimuli that have not been present in the training set, or the degree to which learning one stimulus is transferred to other stimuli.

Question 9

Perceptron

Answer 9

An algorithm for supervised learning

Question 10

Generalisation gradients in pigeons

Answer 10

Come back to

Question 11

Generalisation in one layer network

Answer 11

Networks with local representation cannot generalise

Question 12

Distributed representation

Answer 12

Coarse coding
It may allow the possibility of generalisation.
Representation is more robust; loss of a few neurons will not dramatically degrade the performance of the neural network.

Question 13

In distributed representation, each stimulus is…

Answer 13

…mapped onto more than one node. The mapping can be topographic, but that’s not necessary for the network to perform. A key factor is that the degree of overlap between two stimuli represents their degree of similarity.