BB TB10

?
What did Hubel and Wiesel find?
Cells responding to visual field, edges, specific directions of orientations can be used to create graded tuning curves
1 of 121
What are tuning curves?
As well as sensory domains can be used for cognition and motor
2 of 121
How do you think about only one thing at any given time?
Marilyn Monroe concept
3 of 121
Why did people start thinking there might be one cell for every concept?
Jennifer Aniston study
4 of 121
What does the simpsons study highlight?
Not just seen concepts but thoughts too se concept includes a representation of all elements and is not just a response code
5 of 121
What did Waydo et al (2006) find which suggests there's not just one cell for every concept?
A given neuron fires to 0.54% of all possible stimuli so represents at least 50-150 different concepts
6 of 121
Why is a grandmother cell hypothesis unlikely?
There are many neurons for each concept and many concepts per neuron
7 of 121
How can neural firing be considered ambiguous?
There isn't all-or-nothing firing from a neuron so it's not like a true detector. In fact there is graded firing which explains graded tuning curves.
8 of 121
What are the benefits of having neurons that are broadly tuned?
This can reflect uncertainty which means firing patterns can be generalised across different stimuli and this is useful for forming concepts.
9 of 121
Why is the fact that output firing from a single neuron is ambiguous a bad thing in terms of identifying the stimuli the neuron is tuned for?
The tuning curve generated from a single neuron is symmetrical which means a stimulus with a high valence can create the same output as a different stimuli with low valence.
10 of 121
What happens when there are tuning curves for two neurons in a single population?
The neurons will have slightly different tuning curves which allows the preffered value of each neuron to be disambiguated
11 of 121
When there is coarse coding there is similarity in neural outputs in a single population code. What is good about this?
Similar processing is good for generalisation and this allows concepts to be formed for multiple stimuli
12 of 121
When there is coarse coding there is similarity in neural outputs in a single population code. What is bad about this?
There can be overgeneralisation which means specific peculiarities in the stimuuli that are necessary for firing can be missed. In other words it can be difficult to disambiguate patterns that require different processing.
13 of 121
How do you get accurate information about the necessary input for observed output?
By combining the activity of a whole population of neurons. The preferred weights for each neuron are combined and the preferred values are decoded by neuroscientists allowing predictions to be made about which stimulus the neuron responds to.
14 of 121
What does population coding of different wvelengths by 3 cones allow the brain to do?
Give rise to whole spectrum of colours including those not in the rainbow such as magenta which don't correspond to any single wavelength.
15 of 121
How does metamer perception come about?
Population coding is limited to a narrow slice of the worlds's wavelengths so colours that correspond to a single wavelength of light can also be perceived from mixtures of 2 different wavelengths
16 of 121
What is a tetrochromat?
Affects 5-10% of all women. This is someone with 4 photoreceptors who can perceive a wider spectrum of wavelengths
17 of 121
When the preferred value of a receptor is red and of another receptor is blue then the average wavelength is green. Why then is magenta perceived?
The brain processes patterns of activity across populations rather than decoding the preferred value of single neurons. This is because the brain is able to decode one code by recoding it into another pattern.
18 of 121
What is the sparsest idea of a cell you can imagine?
A grandmother cell but this is thought to not exist
19 of 121
A sparse population code is pretty selective. Why is it not very efficient?
More neurons are required than dense population codes which means more space is required even though it can have a simple architecture.
20 of 121
In what way can a sparse code be thought to be efficient?
Each cell fires infrequently because there are so many cells in total so less energy is used overall
21 of 121
Why is it impossible to get completely distinct codes even in sparse coding and why might the overlap that results be a good thing?
All neurons in a given population respond to similar stimuli so will have similar firing rates. This is good for generalisation (but not to the extent that dense coding has)
22 of 121
Why is sparse population coding better than dense for recognition?
Because there is less overlap between the codes so it is resistant for interference from nearby cells and so is good for noticing differences in stimuli
23 of 121
Why would it be benefical to have sparse neural coding in the hippocamus?
Sparse coding allows rapid learning and encoded stimuli can remain separate from eachother
24 of 121
What is the main point about sparse population coding?
It allows small populations of neurons to carry information about a single event.
25 of 121
In dense population coding all the neurons in a population will be active at the same time. What are the advantages of dense codes?
There is more overlap than sparse codes so it is much better for encoding similarities between stimuli and this is because each neuron in the population has a broad tuning curve.
26 of 121
What are the disadvantages of dense population codes?
They are prone to interference and may reply complex neural architecture with complex learning rules.
27 of 121
What kind of population code exists in V4?
Coding for curvature because V4 encodes shapes.
28 of 121
What does single cell recording from V4 in monkeys viewing shapes show?
There are individual neurons representing different regions of a shape's outline so can decode cell's output to see which kind of curvature the cell is tuned to. When combine tuning curves able to predict the stimulus being viewed to understand V4
29 of 121
If you look at the pattern of activity during preparation and execution of a single cell in motor cortex you can get a tuning curve for action. What happens when you decode the population?
Can see the preferred value of a population by adding up the weighted averages and finding the average population vector to find out the direction the stimulus was moving in
30 of 121
Face cells are located in the inferior frontal cortex, what can decoding population codes here do?
Allow us to predict which face a monkey is looking at
31 of 121
What happens when population codes of place cells are decoded?
Can predict where in space the animal is. Population codes of head direction cells can then tell you which direction the animal is moving in.
32 of 121
What are dendritic arborisations?
Branching patterns
33 of 121
What is the presynaptic terminal of an axon called?
A synaptic bouton
34 of 121
What are the 3 features of a parallel distributed processing model?
Input layer, hidden layers, output layer
35 of 121
In what way is learning said to be unsupervised?
During hebbian learning the connection strength of 2 units is automatically increased whenever units are activated simultaneously?
36 of 121
What does the concept of hebbian learning fit well with?
Evidence that knowledge is stored in distributed regions so local damage leads to degradation but not abolition
37 of 121
What is it called when local damage leads to degradation but not abolition?
Graceful degradation
38 of 121
What happens in single unit electrical recording?
Record extracellularly by measuring voltage as a function of time
39 of 121
Whatare the features of a peristimulus histogram?
A different row for each trial with spikes accross trials forming a histogram and these spikes are time-locked to the stimuli
40 of 121
What is the alternative to obtaining a perstimulus histogram and what are it's features?
Neuronal tuning curves. These define the selective sensitivity of cells to certain values, relative to other cells
41 of 121
What problem to multielectrode recording arrays overcome?
Using only one neuron because a single electrode can have multiple recording points along its length
42 of 121
What is retinotopy?
Mapping topographical representations of the visual field in the eye
43 of 121
What is somatropy?
Mapping somatosensory cortex in relation to body parts
44 of 121
What can the action of a stimulus on neural receptors NOT do?
Cannot represent features of the real world. Particularly because neuroscientists are just representing stimulation in 2D images
45 of 121
What did Hubel and Wiesel find out in terms of plasticity?
The cat cortex has little plasticity
46 of 121
What has been found out about human cortex plasticity?
Cataract can result in permanent loss of sight
47 of 121
What have Merzenich found out?
They amputated the digit of an adult monkey an recorded from corresponding cortex. Activity in cortex increased in response to touching neighbouring fingers which shows cortex does have plasticity
48 of 121
What happens when you repeatedly use one part of the body?
Corresponding cortex size increases, i;e musicians and right hemisphere
49 of 121
What happens when part of the retina is destroyed in monkeys?
The region that is normally responsive to this part of the retina now responds to neighbouring parts of the retina showing the cortex does have some plasticity
50 of 121
What evidence is there that cortical plasticity of adults is limited?
Some people have permanent deficits and some people are less able to learn new skills after brain damage
51 of 121
What is represented in motor maps?
Movements not muscles
52 of 121
What does the firing rate of the motor cortex relate to ?
Changes in muscle force
53 of 121
How do we know activation of the motor cortex is not about creating movements directly?
Neurons discharge before actions are initiated
54 of 121
What do higher motor areas like primary motor cortex and frontal eye fields project to?
Brainstem reticular to lower level circuits to produce coordinated movements
55 of 121
What cannot be predicted from single cell recording?
Direction and amplitude of movement
56 of 121
What are did sparks do?
Averaged activity accross thousands of neurons. This is used to compute the desired movement of a vector.
57 of 121
What controls eye movements?
Eye movements are controlled by higher centres: superior colliculus, which contains a topographic map of eye movements. This sends top-down information to local circuits in the brainstem which coordinated movements.
58 of 121
How did Sparks study eye movements?
Inactivated parts of the superior colliculus in monkeys who had been trained to gaze in direction of juice. Found gaze became biased away from the vector preferred by the damaged neuron. This suggests movements are specified by weighted averages.
59 of 121
What did Georgopolous find out?
Compared vectors from primary motor cortex of a monkey trained to move a stick towards the light. Activity increased well before meovement was initiated showing primary motor cortex encodes planned movements.
60 of 121
When do place cells activate?
When in preffered spatial location in local environment.
61 of 121
What does cognitive neuroscience give us an understanding of?
How information is represented and how these representations interact
62 of 121
How might neurons in V1 represent the orientations of lines and hence letter shapes?
Respond to input (features and feature combos) from feature detectors in retinal cells respondng to simuli edges. When output of many neurons responding to different line orientations are combined you can work out letter shapes
63 of 121
How are facial features represented to recognise emotions?
Response of many cells responding to contrast are combined and then decision neurons select which expression is present based on the pattern of contrast decoded
64 of 121
How would you recognise a face as showing trustworthiness?
Delineate the facial features and compare them to your understanding of social traits. Competition between different facial representations can give a function of dominance.
65 of 121
What happens as number of units in a connectionist model increases?
Number of layers increases
66 of 121
Why are some tasks more complex than others?
Though some outputs are very similar they can have very different inputs and in this instance it will be difficult to find an accurate set of connections
67 of 121
What does a neural network depend on?
The total strength of connections and this is the sum of inhibitory and excitatory connections
68 of 121
What are the 3 features of neurons in PDP models?
Each is modelled as a node/unit without a physical structure, the activation refers to the firing rate of the unit and the total strength of all connection from one unit to another is called the weight
69 of 121
How do you calculate total input (Function)?
activation x weight
70 of 121
Systematic changes made to weights in response to input patterns is based on what?
Hebbian's law
71 of 121
What is the bottom line of supervised learning?
Output pattern is compared to a correct answer so weights are adjusted accordingly to reduce future discrepencies.
72 of 121
Why is output initially different to desired output in supervised learning?
Because connection weights are random
73 of 121
When connection weights are random, what will happen to a canary neuron when a budgie is seen>?
The canary neuron fires too much
74 of 121
What happens in supervised learning when some neurons are tuned to "spots under chin" when repeatedly see birds?
You learn to associate "spots under chin" with budgie so connection weight strength for these tuned neurons increases and connection weights to canary neurons are decresed so these neurons become inhibited and fire less
75 of 121
What happens in supervised learning as connection discrepancy between desired and observed output decreases?
When this is repeated for many input patterns, there will be no error eventually
76 of 121
What does a neural network simulator do?
Connection strengths to feature detectors are learnt by a network so that over time units responding to irrelevant features become inhibited.
77 of 121
What is the main feature of unsupervised learning?
You don't need to know what the correct pattern of activity for an output is
78 of 121
What kind of learning does unsupervised learning explain?
Skills that are picked up without explicit teaching
79 of 121
Why is unsupervised learning more biologically plausible?
Neuronal connections which fire simultaneously increase in strength and this fits with what we know about synapses and memory.
80 of 121
What is one example of unsupervised learning and how does it work?
Competition learning. This is where output neurons compete and those with more regular concurrent activtivations win out whereal other connections are inhibited. This is lateral inhibition
81 of 121
What happens as a result of concurrent connections winning out?
When concurrent input pattern happens again the time taken for the output neuron to win is faster so system becomes more efficient
82 of 121
What happens in competitive learning when there are similar input patterns?
The same output neurons activate
83 of 121
A weight limit is put on synaptic strength of the connections that win out. What does this ensure?
That each output node has a fair chance of responding to each input pattern. This prevents the same neurons from firing again and again while others never fire.
84 of 121
The total sum of squares of total connection weights for each output pattern has a fixed upper limit in competitive learning, how does this lead to inhibition?
When some connections are strenghtened they can't continue to be strengthened indefinitely so to ensure there's always enough disparity with inactive connections, these connections are inhibited.
85 of 121
Why will weight patterns represented as vectors always be drawn the same length?
Because output units have a fixed upper limit.
86 of 121
What are V1 neurons tuned to?
Retinal location, orientation, occular dominance, motion direction and colour
87 of 121
Hubel and Wiesel studied individual neurons, what has been done more recently?
Have used many neurons using crude optical imaging and voltage sensitive dyes. This has found V1 has very systematic pinwheels
88 of 121
What has been found throughout recordings from the visual field?
Similar retinotopic preference with each part of the visual cortex encoding a different part of the visual field.
89 of 121
What has retinotopy found?
cels recorded from same electrode have receptive fields in the same part of the visual field showing theres systematic mapping between areas of visual field and areas of cortex.
90 of 121
What did Inouyes (1907) do?
Took advantage of bullet lesions in V1 and created a very large scale retinotopic map
91 of 121
What has been done since Inouyes?
fMRI on heakthy adults shows retinotopic map. Cn see general organisation of whole visual system doing this but unfortunately can't pick up on structures on the scale of pinwheels (below 1mm)
92 of 121
How is the tonotopic map arranged?
Each hemisphere has an outer area for high frequencies, then low freqiuencies then inner area also for high frequencies
93 of 121
What did Penfield do and find out?
Mapped out the somatosensory cortex by systematically stimulating the brain surface of brain surgery patients. Did motor mapping then and found motor cotex lies right next to somatosensory cortex.
94 of 121
Is it just primary cortex where areas doing similar functions are systematically arranged near eachother?
No, has been found in association cortex too. It is harder to find maps here though because these neurons don't respond to simple variables and often don't know which variables determine their response
95 of 121
What kind of cells have been found in the hippocampus at least?
Place cells. Those with small place fields reside in dorsal part and those with large place fields reside in ventral area which suggests hippocampus has maps of spatial scales. Grid cells have been found here too which give information about distance
96 of 121
What is the general principle of cortical magnification?
The more processing is required, the greater allocation that of cortical space that kind of processing is given.
97 of 121
According to cortical magnification, will the fovea have a bigger or smaller area of V1 cortex devoted to it than periphery?
Receptors are more densely packed in the fovea than in periphery so the fovea requires more processing therefore has a larger proportion of V1 devoted to it
98 of 121
What kind of processing will be given higher cortical magnification in the motor cortex?
Regions necessary for finer motor control
99 of 121
What kind of processing will need more cortical magnification in the somatosensory cortex?
Regions that process areas of body surface with more densely packed receptors
100 of 121
Differentiate the plasticity of different musicians?
String players have unilateral cortex increase whereas pianists have bilateral cortex increase
101 of 121
What are self organising maps?
Models of learning that reflect map organisations
102 of 121
Which researchers were the first to propose self organising maps?
Von der malsberg and willshaw (1972) and Kohonen (1982)
103 of 121
What do we use today to reflect competitive learning?
Modelling
104 of 121
What does activation of neurons in the centre of a population of neurons do?
Neighbouring neurons in the population centre are facilitated but neuros in the surround of the neighbourhood are inhibited. This is lateral inhibition of neighbouring neurons/populations
105 of 121
What happens during lateral inhibition as input continues to be presented?
Weights become more disparate because the winning centre of the population becomes even more positive and the surround more negative
106 of 121
What is the outcome of lateral inhibition in competitive learning?
Physically adjacent units come to represent similar input patterns! This results in local neural changes which explains why neurons with similar functions are arranged systematically near to eachother
107 of 121
What is a term that can explain self-organising maps?
Unsupervised learning
108 of 121
What happens to neural tuning when self organising maps are made?
Eventually tuning spreads evenly across the space of an input
109 of 121
How does unsupervised learning explain cortical magnification?
Simipar input patterns occuring frequently are allocated more space in a cognitive map, whereas infrequent patterns are allocated less space
110 of 121
How do you tell if there is systematic mapping of different kinds of features across a self organising map?
Loook at tuning patters for the mapping of all stimuli
111 of 121
What support does the fruit paradigm particularly give to self organising maps?
The map for fruit contains only features of the stimuli that were presented, so there is no representation for the colour purple because this was never presented in the stimuli.
112 of 121
What happened in Georgopoulos's (1995) study?
Monkey had to move a lever to a circle where a light flashed to receive food. Action potentials of a single motor neuron were recorded and plotted on a rasta plot. This was used to identify which direction of movement the motor neuron was tuned to.
113 of 121
What did Georgopoulous (1995) find?
A high response on the rasta plot for movement in a downwards direction. Motor cortex codes direction of movement and not end location of a moving limb.
114 of 121
What did Georgopolous (1995) find out about movement planning?
The motor neuron responded after stimulus onset but before movement was initiated so motor neurons play a role in planning movements
115 of 121
When looking at tuning curves, what makes it clear that activity from single neurons is very ambiguous?
When the neuron fires at half of it's maximum firing rate a tuning curve for movement direction shows the movement could be in direction A or the opposite to direction A an
116 of 121
What is a population vector?
The average of all tuning curves for action direction
117 of 121
What has Georgopolous' study led to?
Interface devices for prosthetics (Brain-Machine Interface...BMI)
118 of 121
How do you develop interface devices by finding input patterns for individual neurons?
Neural signals can give rise to movements without any muscle feedback so record thought processed for controlling voluntary action which gives signalling for crude movements. Microelectrodes give finer analysis than EEG. Use to code a p. limb.
119 of 121
What happens when a BMI is attached to someone?
Person just has to think about moving their limb and the neural input will match the coding allowing the robot limb to move independent of any muscle activity.
120 of 121
Schwartz pioneered robot prosthetics on what?
On neurons who used it to pick up peanuts
121 of 121

Other cards in this set

Card 2

Front

What are tuning curves?

Back

As well as sensory domains can be used for cognition and motor

Card 3

Front

How do you think about only one thing at any given time?

Back

Preview of the front of card 3

Card 4

Front

Why did people start thinking there might be one cell for every concept?

Back

Preview of the front of card 4

Card 5

Front

What does the simpsons study highlight?

Back

Preview of the front of card 5
View more cards

Comments

No comments have yet been made

Similar Psychology resources:

See all Psychology resources »See all Principles of neural representation resources »