Computational Neuroscience 3

Question 1

Function of the hidden layer in a multilayer network

Answer 1

Contains the distributed representation

Question 2

Training for similar stimuli …

Answer 2

…reinforces the original one i.e training a network trained for yellow for orange reinforces yellow, in the network and in animals thus distributed representation is supported in vivo

Question 3

Backpropagation

Answer 3

A mechanism in which weights are adjusted starting from those in the outer layers back to those in the input layers. It is a generalisation of the Widrow-Hoff rule.
Now the standard algorithm for training networks at WH rule can’t train MLNs

Question 4

Unsupervised learning

Answer 4

Learning by simple exposure, e.g. the autoassociative networks in the hippocampus

Question 5

Unsupervised learning requires…

Answer 5

…a different neuronal architecture and/or a different rule

Question 6

Animals with hippocampal damage…

Answer 6

…cannot learn unsupervised.

Question 7

Network structure compatible with unsupervised learning

Answer 7

CA3

Question 8

Pattern completion

Answer 8

(A) Two different patterns are stored in the network
(B) An incomplete patterns in presented
(C) Because of the synapses linking the nodes in the pattern assembly,
activity spreads.
(D) The pattern is recovered

Question 9

Pattern recognition

Answer 9

(A) Two different patterns are stored in the network
(B) A pattern not stored in the network is presented
(C) Strong synapses spread the activity.
(D) Activity in cells that do not belong to a stored pattern decays, and a
stored patterns becomes active.

Question 10

The Hopfield Model

model for auto-associative memory

Answer 10

Every neurone synapses onto every other neurone.
Weights are set according to a Hebbian rule.
Each vector (pattern) is represented as the desired activation state of the network for that pattern.

Question 11

Behaviour of the Hopfield Model

Answer 11

The model has a temporal behaviour (dynamics) such that it can evolve over time.
The ‘memories’ stored are stable attractors of the dynamics.
If the model is initialised closed to a memory (with a partial pattern) it will evolve towards the closest attractor. Thus, retrieving the stored memory by association.
The model always tries to look for the closest trough in the energy function.

Question 12

When presented with an incomplete or distorted pattern, the Hopfield network…

Answer 12

…evolves to the closest attractor (stored memory)

Question 13

The hippocampus and autoassociation

Answer 13

Area CA3 receives input from EC (highly processed information from sensory cortex). It also receives input from DG, which in turn receives its input from EC. Then CA3 sends information to CA1 and Fornix, and then from CA1 to S.
The connections from DG to CA3 are called Mossy Fibers. They make sparse and powerful synapses onto CA3 neurons.
CA3 has a high degree of internal recurrency and random connectivity.
Each CA3 neuron in the rat receives 4000 synapses from EC and 12000 from CA3 itself. That is a level of recurrency orders of magnitude higher than in other brain areas.
The connections between CA3 neurons are modified by LTP very fast (minutes or hours)

Question 14

Difference between sensory system architecture and hippocampal architecture

Answer 14

Sensory systems which have a direction of information flow, hippocampal architecture is heavily recurrent

Question 15

Storage

Answer 15

(A) Synapses from Mossy Fibers elicit strong activity in a cell assembly (B) Synapses within the assembly strengthen via LTP (Hebbian learning)

Question 16

Retrieval

Answer 16

C) Weaker synapses from EC induce low level activity in a cell population
(D) Because of the strong synapses the previously stored pattern is recovered.

Question 17

It has been proposed that mossy fibres are…

Answer 17

…responsible for the strong activation CA3 ‘nodes’ –> undergo Hebbian learning thus LTP, so could Mossy fibres be responsible for storage?

Question 18

Limitation of autoassociative networks

Answer 18

Catastrophic interference occurs when the number or stored patterns increases
(occurs in the Hopfield model)

Question 19

Marr proposed that…

Answer 19

CA3 can only hold memories for a limited period of time after which the memories are transferred to the neocortex