Neural module

Question 1

How can we compare the spiking activity responses to different stimuli?

Answer 1

1) Grand-average method:
– one grand average per neuron
– perform standard (non)parametric test
2) Single-trial method:
– Each trial will yield a response
– Perform a repeated measurement (random effects) test

(Both methods are valid, and often equivalent, although the latter can be more powerful;
Rule of thumb: use the simplest available method)

Question 2

What is the Coefficient of variation (in the context of PSTH) and how do you compute it?

Answer 2

A measure of response reliability(/how stereotypical response is)
1. Compute response parameter for each trial (single-trial PSTH)
2. Build probability distribution of responses
3. CV = sigma/mu

Question 3

What is the response period in PSTH and what variables can we extract from it?

Answer 3

Response period: interval after the stimulus where activity is
significantly greater than in the baseline (> mean + 2 s.d.)
– Average/total activity in response period (option: –
mean(baseline))
– Max activity in response period (option: – mean(baseline))
– Peak latency from stimulus onset
– Onset latency from stimulus onset (delay of first significant response bin)

Question 4

What are spike trains and what are the two major ways of computing firing rates based on them?

Answer 4

A spike train is a series of timestamps representing the times at which a neuron fires an action potential (spike).

Discrete-time firing rates: time is divided into equal-length bins, and the firing rate is calculated for each bin as the number of spikes that occur within that bin divided by the bin width.

Sliding window: a window of fixed length is “slid” across the spike train, and the firing rate is calculated for each position of the window based on the spikes within it ((spike train is convolved with a window function). Unlike discrete bins, the window overlaps as it slides, providing a smoother estimate. There are different types e.g rectangle window, gaussian window, alpha function

Question 5

Decoding vs. encoding (in the context of neural spikes)

Answer 5

Encoding: how neurons codify information about sensory stimuli.
– given a sensory stimulus, what is its neural representation in terms of
spiking activity?
– what is the probability of observing a certain spiking pattern, given a
sensory stimulus?
Decoding: what was the stimulus, given a certain pattern of neuronal activity?
– probability of a stimulus, given a pattern of activity

Question 6

What is the simplest type of classifier?

Answer 6

Linear classifier: Identify the line that best separate the two classes

Question 7

1) How do we identify which line best separates two classes and (2) how can we automatically and optimally do this?

Answer 7

1. Minimize the error in the classifier (cost or loss function)
2. Neural networks

Question 8

How is learning achieved in a perceptron?

Answer 8

1. Fix a training set of M samples

Question 9

What is the architecture of an MLP?

Answer 9

• All the neurons of a layer are connected to all the neurons of the next layer.
• There are no connections between neurons in the same layer and between
  non adjacent layers.
  Multi-layer perceptrons can separate any convex region

Question 10

How do we compute the error for hidden layers?

Answer 10

backpropagation algorithm.

Question 11

Which rule are weights varied according to?

Answer 11

Learning rule/optimization algorithm, e g: Gradient descent rule

Question 12

What is an epoch in the context of training an MLP and what steps does it entail?

Answer 12

A pass of the entire training set.
1. Forward Propagation:
Input data is passed through the network, layer by layer, applying weights, biases, and activation functions.
The network computes predictions (output).

2. Loss Computation
   Compare the predicted outputs with the actual labels using a loss function appropriate for the task:
   * Mean Squared Error (MSE) for regression.
   * Cross-Entropy Loss for classification.
3. Back propagation:
   Compute the gradient of the loss function with respect to the weights and biases using the chain rule of calculus (backpropagation algorithm).
   Gradients are propagated backward through the network from the output layer to the input layer.
4. update weights (how often this is done depends on training strategy)
   * based on learning rule/optimization algorithm (e.g gradient descent)

Question 13

How do we evaluate the capability of an MLP to generalize the examples of the training set?

Answer 13

By testing it on the Validation set (that sounds to me like its the same thing as a test set but idk)

Question 14

Three things that are important to ensure good training, generalization and no overfitting

Answer 14

• Network architecture (number of layers/neurons)
• learning rate
• a good
  choice of TS

Question 15

What is cross validation?

Answer 15

a procedure to prevent a classifier from achieving optimal
performance only on a specific validation/test set, and to test a classifier’s
performance against the whole available dataset

Question 16

How are spikes plotted in a raster plot?

Answer 16

each onset-aligned spike is plotted as a dot, one trial per row.

Question 17

What is dPCA

Answer 17

dPCA (demixed PCA):
dPCA extends PCA by “demixing” the variance in neural activity into components that can be linked to specific variables of interest.
It separates the variability in neural data into distinct contributions (e.g., stimulus-driven, motor-response-driven, decision-related components), making the analysis more interpretable.
Each dPC (demixed principal component) is explicitly aligned with a task-related variable or factor.

Question 18

Spike sorting steps

Answer 18

spike sorting steps:
* Spike extraction
* Waveforms alignment
* PCA
(in the plot they have also done t-SNE but not in the description outlining the steps).
* Clustering
(5. Manual curation)

Question 19

What is PCA a technique for and what is the goal of it?

Answer 19

PCA is a technique for identifying the most relevant variables, i.e. able to explain most of the variance in the dataset
Uses:
* Data compression (i assume that is dimensionality reduction??)
* feature extraction
* reduced computational burden on clustering algorithm
* removal of “noisy” (non informative) components from data

The goal of PCA is:
1-Find the vectros describing the pc1, pc2, etc. directions
2-Project the data points on the new coordinate system

Question 20

Explain what type of algorithm k-means is, how it works and what the pros/cons are?

Answer 20

K-means algorithm
* Partitional, model-independent, exclusive clustering algorithm
* Commonly based on euclidean distance
* The cost-function to minimize is the within-cluster sum of squares
* Clusters are found via an iterative algorithm, with 3 steps:
1. Initialization: Cluster number and centroids are initiated by the user (centroids usually randomly chosen)
2. Assignment: Data points are assigned to closest centroid
3. Update: centroid coordinates are calculated again, as the center of mass of each cluster

Pros: - fast, easily interpretable. Cons: different solutions depending on initialization, may have shapes that do not correspond to the actual distribution of data

Question 21

What is the purpose of PSTH and what are the steps?

Answer 21

1. Select the number of trials
2. Select bin size (1 ms: time code, >25 ms: rate code)
3. Compute average number of spikes in each bin
4. Divide by bin width to compute firing rate