Lecture 10: Networks

Question 1

What are principles of neural network processing?

Answer 1

1) distributed representations
2) connectivity
3) excitation/inhibition

Question 2

neural representation

Answer 2

ensemble of neurons that fire in synchrony in relation to stimulus

Question 3

grandmother cell concept

Answer 3

cells responding consistently to specific face

Question 4

Why grandmother concept is questionable?

Answer 4

because there is no single cell responding to grandmother -> there is ensemble of cells responding to the face
also, those cells are not specifically tuned to the grandmother face -> they also were activated by different faces
suggesting overlap between assemblies of cells

Question 5

Why nervous system uses assembly of representation? = distributed representations

Answer 5

1) increasing capacity of nervous system to represent things
2) more resistance to damage
3) more robust to noise
4) increase in dynamic, flexible behavior of networks
5) allows competitive behavior

Question 6

What is principle of long-term potentiation?

Answer 6

'’neurons that fire together, wire together’’ - when activity in one cell repeatedly elicits action potentials in second cell, synaptic strength is potentiated

Question 7

What is long term depression?

Answer 7

activity-dependent reduction in the efficacy of neuronal synapses lasting hours or longer following a long patterned stimulus
one of several processes that serves to selectively weaken specific synapses in order to make constructive use of synaptic strengthening caused by LTP
this is necessary because, if allowed to continue increasing in strength, synapses would ultimately reach a ceiling level of efficiency, which would inhibit the encoding of new information
enables isolating the pattern from the background

Question 8

long term neuronal representation

Answer 8

relatively strong synapses between neurons of the ensemble
relative to what?
-> to other synapses/representations in synaptic matrix
-> especially for competing (representionally overlapping) information

Question 9

what happens when you retrieve information?

Answer 9

cue (relating to previously encountered stimuli) leads to partial activation of relatively strengthened synapse between neurons - continuous cueing will lead to pattern completion (the rest of pattern gets activated)

Question 10

What are 2 types of plasticity?

Answer 10

1) autoassociation = local synaptic strengthening = with one layer - within one representation
2) heteroassociation = representation in layer 1 triggering representation in layer 2 - synaptic strengthening between different representations

Question 11

What is example of importance of connectivity in information processing?

Answer 11

Visual processing
retina - light sensitive photoreceptors -> rodes and cones -> ganglion cells fire action potentials to optic nerve which (amongst others) project to thalamus
receptive fields of ganglion cells are circular
however, in V1 receptive fields are line-shaped (selective responses to line orientations)
why? because one V1 cell gets input from set of LGN cells which get input from ganglion cells receptive fields oriented in line

Question 12

What are general principles of connectivity patterns?

Answer 12

1) divergence
2) convergence
3) point to point (topological) connectivity

Question 13

divergence

Answer 13

spread of information from one source cell to multiple target cells

Question 14

convergence

Answer 14

compression (combination) of information from multiple source cells into 1 target cell

Question 15

point to point (topological) connectivity

Answer 15

number of source and target neurons is the same, copy information from source to target layer
relation to environmental topology typically diminishes as we ascend the neural hierarchy (lost at higher levels - associations areas)

Question 16

What are types of divergence?

Answer 16

1) from source area to multiple targer areas
example: infromation from retinal cells diverges over what (temporal lobe) and where (parietal lobe) pathways

2) from source to one target area
example: input from layer 1 is spread over very different cells in layer 2 (easier to combine incoming information) - often in association networks

Question 17

What are types of convergence?

Answer 17

1) with one modality
example: visual processing (LGN -> V1 -> V2 -> V3 -> V4) - from light patches to representation of the whole receptive field

2) across modalities
example: episodic memory system (hippocampus combining input from entire sensory field)

Question 18

What are differences in connectivity patterns across processing hierarchy?

Answer 18

at lower levels: more point-to-point
because lower levels represent patterns of environment which are pretty stable (V1: statistics of environment with regards to orientation)

at higher levels: association areas -> more fanning -> more higher degree divergence and also convergence
representations of more variable aspects of environment
more plasticity

Question 19

What is feedforward inhibition?

Answer 19

as layer 1 sends excitatory signal to layer 2
the same signal activates GABA neuron in layer 2!
proportional inhibition activity to source excitation
allows network to handle input of variable strength

Question 20

What is feedback inhibition?

Answer 20

layer 1 sends excitatory signal to layer 2
excited neuron in layer 2 will activate interneuron which will send BACK inhibitory signal
proportional to activity in target cell
limits excitation time - neuron gets excitated, but inhibition follows very quickly
ensuring no association patterns in level 2 will get activated
contributed to gamma oscillations

Question 21

What is control of feedforward and feedback inhibition?

Answer 21

level of firing in a layer
level of propagation

Question 22

How level of firing in a layer enables inhibition control?

Answer 22

it relates to pattern size
for example: hippcampus has high level of inhibition, one inhibitory neron can inhibit many, many cells, small pattern size

Question 23

How level of propagation enables inhibitory control?

Answer 23

too little inhibition = exploding excitation and multiplication of noise
too much inhibition = propagation dies out across layers and becomes highly conservative

Question 24

What is neural representation of outside world? (continuous vs discrete)

Answer 24

Most likely snapshots (discrete).
Why? because inhibition in a layer is total = as representation is shut down by inhibition, there is room for new representation to emerge