Brainscape's Knowledge GenomeTM

Browse over 1 million classes created by top students, professors, publishers, and experts.


See full index

ML1 > ML2 > Flashcards

ML2 Flashcards

1
Q

Why study biological networks?

A
  1. Inspiration for many AI networks, including DCNNs
  2. AI networks simplify processes involved for efficiency
    - AI deep networks not yet as advanced as human brain
    - Where the goal is to imitate human behaviour, following the neural mechanisms more closely may help
  3. Major link between AI and other sciences
    - Links AI to biological sciences, not just math/CS
    - Potential to link social sciences to biological sciences
  4. Leading model of neural computations.
How well did you know this?
1
Not at all
2
3
4
5
Perfectly
2
Q

What does the cell membrane in a neuron do?

A

It is specialised for performing simple computations using electrical activity.

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
3
Q

Explain the process of how the cell membrane in the neuron works.

A
  1. The dendrites integrating electrical signals coming in from other neurons that it is connected to.
  2. These connections, or synapses, between dendrites and other neurons vary in strength, and change strength depending on past activity, so synaptic strengths are the biological equivalent of artificial neural network weights.
  3. We will see that the biological equivalent of a convolutional filter is a tree of dendrites working together to sample information from other neurons.
  4. If these combined incoming electrical signals are strong enough to pass a THRESHOLD, the neuron will then send an electrical impulse down the axon.
  5. On reaching the axon terminals, this is passed on to other cells, in the next LAYER of neurons.
How well did you know this?
1
Not at all
2
3
4
5
Perfectly
4
Q

Why is the cell membrane really important?

A

This is where the electrical signals take place.

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
5
Q

What is the basic computational component of a biological neuron?

A

A protein called an ion channel

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
6
Q

The ion channel is a passive mechanism, what does it mean?

A

It does not pump the ions against their concentration gradient, it only opens to allow ions to diffuse down the concentration gradient or closes to prevent this.

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
7
Q

What is the resting potential?

A

-70 mV

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
8
Q

Name the 1 important way an ion channel can open.

A
  1. It can open because a neurotransmitter binds to the ion channel. In that case, the ion channel is a ‘receptor’ for the neurotransmitter.
    When the neurotransmitter binds, the ion channel protein changes shape, opening the ion channel for ions to cross the membrane.
    Another neuron firing causes the neurotransmitter release into a synapse, the gap between two linked neurons.
    So the neurotransmitter is a signal released by activation of one layer of a neural network (the presynaptic neuron) and causes activation of the next layer of the neural network (the postsynaptic neuron).

Here we will look at excitation (prikkeling) of the postsynaptic neuron by binding of glutamate, and inhibition by binding of gamma-amino butyric acid (GABA). Using both excitation and inhibition allows connections with both positive and negative weights, as convolutional filters have.

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
9
Q

Name another important way an ion channel can open.

A
  1. An ion channel can open due to a change in membrane voltage. This is particularly important because opening the ion channel also changes the membrane voltage.
    - So changing membrane voltage can lead to further changes in membrane voltage.
    - In this way, the voltage-gated ion channel acts very much like an electronic transistor, that is an electrical switch that opens and closes because of an electrical input. Miniaturised transistor circuits are the basis of all computer processors.
How well did you know this?
1
Not at all
2
3
4
5
Perfectly
10
Q

The threshold for activation of voltage-gated sodium channels is the biological equivalent of …

A

… the threshold/rectification operation we saw in artificial networks.

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
11
Q

What is the threshold in biological neurons?

A

-15 mV

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
12
Q

Does the firing rate (output) of the neuron reach a maximum firing rate?

A

yes

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
13
Q

Do we have both positive and negative inputs on to the same postsynaptic neuron?

A

yes. These can be from a range of different places, and from a number of different neurons. However, there is a limited spatial distribution of inputs because the dendritic tree has a limited size.

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
14
Q

The dendritic tree has a limited size. Which may sound very familiar …

A

Because convolutional network filters imitate this structure: they have positive and negative filters, which are multiplied by the activity on a group of presynaptic units to give the activity of the postsynaptic unit

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
15
Q

Name the 2 neurotransmitters and what they do.

A

Neurotransmitters (e.g. glutamate and GABA released from a pre-synaptic cell can excite (depolarise) or inhibit (hyperpolarise) activity in the post-synaptic neuron
- this relies on ligand-gated (i.e. neurotransmitter activated) ion channels

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
16
Q

If membrane polarisation reaches a threshold, voltage-gated ion channels open

A
  • Results from many excitatory inputs

- Strongly depolarises (fires) the neuron: Action potential

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
17
Q

Once an action potential has been triggered, …

A

it travels from the input end of the neuron, the postsynaptic dendrites, along the length of the neuron’s main fibre, the axon, to provide inputs to the next layer, which is often in a different brain area some distance away.

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
18
Q

The spiking depolarisation at one location spreads to neighbouring locations to push their membrane potential above threshold. Does the depolarisation spread on the axon like a wire?

A

yes

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
19
Q

Weights in biological neurons (True or False)

  • Connection weights in biological systems don’t depend on backpropogation of error
  • We are trained by a supervised process
  • However, there can be adaptive and maladaptive responses/behaviours
A

True
False
True

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
20
Q

We learn mainly by unsupervised process

A
  • To recognise patterns of activity we have seen before

- To learn the statistics of the world we live in

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
21
Q

Hebb’s postulate

A

Cells that fire together, wire together

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
22
Q

Also cells that wire together lie together. This …

A

Reduces required connection length and increases efficiency/speed

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
23
Q

Which common drugs reduce spiking activity by activating GABA receptors and reducing the probability of depolarisation?

A

Alcohol and benzodiazepines

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
24
Q

Filters are …

A

integration of EPSPs and IPSPs across the dendrite tree

25
Q

Threshold …

A

activation of voltage-gated Na+ channels by above-threshold depolarisation by EPSPs

26
Q

Learning mechanism …

A
  • unsupervised: ‘cells that fire together, wire together’ (Hebbian learning)
  • Activity-dependent changes in synapse structure/strength (weights)
27
Q

The eye has four different sensors for light: three types of colour-sensitive cones sensor and one type of rod sensor.

A

Rods don’t carry colour information, but respond quickly and under low light conditions. Their output also gives good information about fast events like moving objects.

Together, these form the input image to the visual processing network, and are essentially already four feature maps.

28
Q

Are the light sensors evenly distributed across the image?

A

No

29
Q

Cone density is much higher in the …

A

central vision (‘where we are looking’)

30
Q

Rods are much denser …

A

away from central vision

31
Q

Which are larger: cones or rods?

A

Rods

32
Q

Because cones are smaller, …

A

they each respond to a smaller area of the image, so give finer vision where we are looking.

33
Q

vision strongly over-emphasises the …

A

centre, where we are looking!!

34
Q

Why over-emphasis the centre?

A

This greatly reduces the computational load on the brain, while still giving high detail in central vision.

35
Q

How does vision processing work?

A

neural representations of image transformations (rather than thinking of this as an image!!)

36
Q

Why is it very important that neighbouring locations are represented next to each other?

A

This is because then the spatial extent of a filter represents a continuous piece of the input image.

37
Q

What is the main reason why the brain maintains spatial relationships at each level of processing?

A

When we consider that the dendrites of the next neuron will sample from this layer, and have a limited extent too.

38
Q

What sensory and mortor areas of the brain have larger, more detailed representations?

A

hands, faces, and tongues: sensitive body parts that perform detailed movements

39
Q

Artificial sensory systems generally don’t use such distorted inputs, aiming instead to process everything in great detail. What are the advantages/disadvantages of this?

A

The main disadvantage is that processing the whole image input in similar detail is computationally intensive. The main advantage is that the cameras that provide the artificial networks don’t need to move to sample the important parts of the image in detail, like our eyes do. Normally, the inputs are static images, so this isn’t an option anyway.

40
Q

Which two responses form separate feature maps for light contrast and dark contrast?

A

The set of on-centre and off-centre

41
Q

What do the first filtering stages in the eye?

A

They analyse the relationships between nearby locations, and simultaneously the relationship between the different colour maps

42
Q

The retina has several stages between the photoreceptor and the ganglion cells of the optic nerve to the brain. The first of …

A

The first of these compares the responses of a group of neighbouring photoreceptor cells, all linked in to its tree of dendrites. Those in the centre of the tree produce EPSPs, while those at the edges of the tree produce IPSPs. So activation at the edges inhibits responses to activation at the centre

43
Q

What produces a stronger response: a point of light at the centre or a field of light co bring bot the centre and the surround?

A

A point of light at the centre.

44
Q

Just like a convolutional filter, there are overlapping copies of this cell throughout the retina. What does this mean?

A

A photoreceptor that falls in the negative zone of one horizontal cell filter, also falls in the positive zone of another.

45
Q

What is the complementary filter recording spatial comparison filters?

A

Operating a parallel, responding to darkness at the centre and suppressed by darkness in the surround. This is called an ‘of’ response, a response to no light. This might sound unnecessary, but features are often darker than the background, like the text on this page.

46
Q

Explain how these cells (retina, photoreceptor, ganglion) respond to different patterns of light and darkness throughout their responsive area.

A

The set of on-centre and off-centre responses effectively form separate feature maps for light contrast and dark contrast.

47
Q

Why is contrast necessary?

A

Because a full field of light or darkness produces no response. More generally, the visual system responds to changes rather than constant inputs. These spatially-specific ‘changes’ can be thought of as ‘features’.

48
Q

What is the area producing the positive response called? And the area suppressing the response?

A

receptive field, suppressive surround

49
Q

What is the term ‘receptive field’ used for in an artificial neural network?

A

the spatial extent of a filter.

50
Q

What does surround suppression do?

A

it normalises the response to light intensity inside the receptive field by the light intensity outside the receptive field

51
Q

What should the light level in the surround do?

A

Normalise the light elven in the receptive field, yielding local contrast.

52
Q

By transforming the representation of brightness in the rods and cones into a representation of contrast, most of the image can be discarded. Which parts?

A

The parts that contain no change

53
Q

The changes in the image are maintained!

A

This keeps all the useful information, but greatly compresses it for transmission down the small optic nerve to the brain.

54
Q

Surround suppression and normalisation closely resembles the normalisation operation of deep networks, and effectively converts light level into contrast level at the first stage. True or false?

A

True

55
Q

However, this mean light level used here is probably not taken over the whole image: there is limitation to the distance that the neuron can be connected to.

A

Furthermore, the mean light level is probably restricted to a single feature map. It’s unclear whether these differences are better or worse than global normalisation. They allow more complex patterns of normalisation, which might be useful if an image has light parts and dark parts, but are limited in spatial extent.

56
Q

Together with spatial comparison, there is also a comparison over the different ….

A

colour cones

57
Q

Name the four feature maps for colours

A

greenness, redness, blueness, and yellowness

58
Q

To detect brightness and darkness, all three cones …

A

activate or inhibit the response

ML1 (3 decks)

Brainscape helps you reach your goals faster, through stronger study habits.
© 2024 Bold Learning Solutions. Terms and Conditions