lecture 8 - normalization

Question 1

What was the early focus of neural circuit research?

Answer 1

Researchers sought canonical circuits—specific physical arrangements of neurons and their interactions.

Question 2

How did the focus shift in neural circuit research?

Answer 2

The focus shifted to computations, where the same computation can be implemented by different circuits across brain areas.

Question 3

What did Marcus propose regarding canonical neural computations?

Answer 3

Marcus proposed a diverse set of computationally distinct building blocks that implement a broad range of elementary, reusable computations.

Question 4

What did Carandini and Heeger define as canonical neural computations?

Answer 4

Carandini and Heeger described canonical neural computations as standard computational modules that apply the same fundamental operations across different contexts.

Question 5

What is the primary goal of studying canonical neural computations?

Answer 5

To identify a computational “core” that underlies diverse functions of the neocortex.

Question 6

What is the consensus about canonical computations?

Answer 6

The emphasis is on identifying fundamental computational principles that apply broadly across brain regions, rather than focusing on specific circuits.

Question 7

What are the three candidates for canonical computations?

Answer 7

1. Receptive fields: Weighted linear summation combining inputs with specific weights.
2. Predictive processing: Bayesian prediction-error propagation.
3. Divisive normalization: Ratio of input-driven and contextual activity.

Question 8

What is the function of receptive fields in canonical computations?

Answer 8

Receptive fields combine inputs with specific weights using weighted linear summation.

Question 9

What is predictive processing in the context of neural computations?

Answer 9

Predictive processing involves Bayesian prediction-error propagation.

Question 10

What is divisive normalization in the context of neural computations?

Answer 10

Divisive normalization is a canonical neural computation for contextual processing

Question 11

What are some phenomena explained by canonical circuits?

Answer 11

1. Persistence of excitation and inhibition: Excitation and inhibition last longer than synaptic delays.
2. Recurrent intracortical amplification: Thalamic input does not provide the major excitation; intracortical excitatory connections amplify input.

Question 12

What role do computational accounts play in neural circuits?

Answer 12

Computational accounts explain the functional purpose of microcircuits, providing insights into their role in complex brain computations.

Question 13

How does divisive normalization work computationally?

Answer 13

• Neuronal responses are the ratio of input-driven (activation pool) and contextual activity of other neurons (normalization pool).

Question 14

What role does context play in divisive normalization?

Answer 14

• Context represents other inputs that are summed and used to divide the original input, providing computational context.
• this way, input → output is influenced by other inputs

Question 15

What nonlinear phenomena in V1 were first explained by divisive normalization?

Answer 15

1. contrast saturation
2. surround suppression

Question 16

In which other situations has divisive normalization been applied?

Answer 16

1. Olfaction
2. Retina
3. Attention
4. Multisensory integration
5. pRFs (population receptive fields).

Question 17

How is divisive normalization conceptually linked to artificial neural networks (ANNs)?

Answer 17

1. Max-pooling: Selecting the maximum value in a neighborhood.
2. Softmax: Scaling outputs to probabilities in classification tasks.

Question 18

What is the generic formulation of divisive normalization?

Answer 18

• The response y is computed as y = [y1(x) + b]/[y2(x)+d]
• here, y1 is the stimulus and y2 is the normalization pool

Question 19

What does divisive normalization suggest about neural activations?

Answer 19

Neural activations are not just input-driven but reflect a ratio of input-driven and contextual activations.

Question 20

What is contrast saturation in V1 neurons?

Answer 20

Neuronal responses increase with contrast but eventually saturate because both y1 and y2 increase. This prevents neurons from endlessly firing as input strength rises.

Question 21

How does center-surround suppression in V1 work?

Answer 21

• Neurons’ responses decrease when there’s increased stimulation in the surround.

1. The normalization pool (neighborhood) is activated more than the activation pool.
2. This increases the denominator y2, reducing the center’s response.

Question 22

How does responsiveness to the center get modulated in surround suppression?

Answer 22

The center’s responsiveness is modulated by the surround, meaning stimulation in the surround influences the response of the center neuron.

Question 23

Why do we want to describe attention through divisive normalization?

Answer 23

Attention modulates how strongly a neuron responds to a stimulus by amplifying or diminishing the response depending on the focus of attention.

Question 24

What are the two primary effects of attention in the brain?

Answer 24

1. contrast gain effect
2. response gain effect

Question 25

What is the contrast gain effect

Answer 25

• Large attention/normalization pool increases contrast gain.
• Attention shifts sensitivity of the neuron to lower contrast levels.
• Neurons start responding at lower contrasts but saturate at the same response percentage as without attention.

Question 26

What is the response gain effect

Answer 26

• Small attention/normalization pool increases response gain by amplifying the attended stimulus.
• The overall response of the neuron increases as contrast increases

Question 27

How does the DN (Divisive Normalization) model of attention explain contrast and response gain?

Answer 27

• by implementing attention as a modulator of the size of the normalization pool

1. large area of attention: creates a large normalization pool that distributes attention across multiple stimuli, leading to contrast gain.
2. small area of attention: creates a small normalization pool that amplifies the attended stimulus, leading to response gain

Question 28

What is population response in attention and divisive normalization?

Answer 28

Population response is the collective neural activity across multiple neurons, accounting for:

1. Stimulus activation
2. Attention gain
3. Divisive normalization effects

Question 29

What mechanism prevents overexcitation of neural responses?

Answer 29

Normalization factors regulate neural responses to prevent overexcitation, representing competition between stimuli in the neural network.

Question 30

What is suppression in signal time-courses?

Answer 30

• Suppression refers to the reduction in the response of a neuron due to inhibitory inputs, often from the surrounding context of the receptive field.

Question 31

How does the Gaussian model explain suppression?

Answer 31

A single Gaussian model captures activations but explains only half the variance.

Additional negative activations (not captured) indicate suppression.

Question 32

What is the Difference of Gaussians (DoG) model?

Answer 32

The DoG model proposes a larger receptive field that subtracts the surround RF from the center RF. It explains negative responses that Gaussian models miss.

Question 33

What is compression in signal time-courses?

Answer 33

• Compression describes the sublinear scaling of neuronal responses as input intensity or size increases.
• Output is compressed at higher inputs

Question 34

How does compression manifest with stimuli?

Answer 34

The response to parts of a stimulus is not summed linearly when showing the full stimulus. Responses grow sublinearly with stimulus size.

Question 35

Which model explains response compression?

Answer 35

The Compressive Spatial Summation (CSS) pRF model explains response compression. It raises the linear response to some power (n < 1).

Question 36

Why should suppression and compression be combined into one model?

Answer 36

• The DoG and CSS models don’t explain each other’s effects

Question 37

Which model explains suppression and compression

Answer 37

• Suppression and compression effects are better explained together in a divisive model

1. Divides the center and surround activation.
2. Outperforms Gaussian, DoG, and CSS models.

Question 38

How are suppression and compression implemented in the pRF model?

Answer 38

1. Activation terms: Driven by the stimulus.
2. Normalization terms: Averages surrounding activity.

Question 39

What ensures the response is zero when there is no stimulus?

Answer 39

The term -b/d ensures the response is 0 when there is no stimulus

Question 40

What is the role of the activation constant b>

Answer 40

• modulates suppression
• When b=0, there is nothing to be suppressed, leading to a linear model.
• As b increases, suppression occurs, and responses can go below baseline (negative flanks).

Question 41

What is the role of the normalization constant d?

Answer 41

• modulates compression
• Large d: Divides by a large value, resulting in a linear response.
• Small d: Leads to a stronger division effect, creating a larger compressive response.

Question 42

What do activation and normalization constants do?

Answer 42

Simulations suggest that activation (b) and normalization (d) constants jointly modulate suppression and compression.

Question 43

How do suppression and compression change from lower to higher levels of the visual cortex?

Answer 43

As you go from lower to higher levels of the visual cortex:

1. b-constant decreases → less suppression.
2. d-constant decreases → more compression.

Question 44

How do the DoG model and CSS model perform across visual cortex levels?

Answer 44

1. The DoG model better explains V1 and V2 (lower areas).
2. The CSS model better explains higher-level visual regions.

Question 45

What is a key free parameter of the normalization pool?

Answer 45

The size of the normalization pool.

Question 46

What is spatial oversaturation in size-response curves?

Answer 46

• Spatial oversaturation occurs when neuron responses increase, peak, and then decrease as stimulus size increases.
• The normalization pool becomes more stimulated than the activation pool, increasing the denominator and decreasing the response.

Question 47

What are the two nonlinearities in the size-response curve?

Answer 47

1. Rise to the peak (Standard Compression): Controlled by the d-parameter.
2. After the peak (Oversaturation): As stimulus size increases, the response further decreases.
   Controlled by the size ratio ϕ-parameter.

• Only the DN model can describe both nonlinearities present in size-response curves.

Question 48

What does the size ratio ϕ represent?

Answer 48

• The ratio of the size (𝜎) of the normalization pool relative to the activation pool.
• strength of spatial oversaturation in visual cortex
• Lower ϕ is less oversaturation

Question 49

How does ϕ change from lower to higher cortical regions

Answer 49

ϕ decreases, so there is less oversaturation

Question 50

Response properties are described by the DN pRF model

Answer 50

Different responses to visual stimuli throughout human cortex

Question 51

What underlies response properties

Answer 51

Activation/normalization constants and pRF size ratio underlie response properties (suppression,
compression, oversaturation).