Collective coherent cortices: Data analysis

Background

In this exercise, we will investigate physiological recordings from visual areas V1 and V4. The experimental setting is comparable to the paradigm used in Attention selectively gates afferent signal transmission to area V4:

Each trial started with a baseline period of 500 ms without visual stimulus. Then two morphing shape sequences were presented for at least 1000 ms. The task for the animal was to attend one of the two stimuli, and to signal the reoccurrence of the initial shape of the attended sequence by releasing a lever.

During the experiment, electrophysiological recordings were done with $N=100$ electrodes (arranged in a 10 by 10 grid) in area V1, and a single electrode in area $V4$ whose receptive field (RF) covered both visual stimuli to a similar extent. The morphing shapes (let’s name the sequences $A$ and $B$) were tagged with independent luminance variations (‘flicker’) for allowing the scientists to track visual information in the neural signals.

Previous work in macaque monkeys established that visual signals are selectively routed in dependence on attention: computing the spectral coherence between V4 LFPs and flicker signals $A$ and $B$ reveals a substantially higher signal content of the attended flicker signal in the physiological data compared to the non-attended flicker signal (see Attention selectively gates afferent signal transmission to area V4). In addition, it has been found that selective signal transfer goes along with enhanced phase synchronization between the corresponding visual areas (see Switching neuronal inputs by differential modulations of gamma-band phase-coherence).

Can these results be reproduced?

One for your classmates is a researcher studying selective processing in the visual system. Your classmate has recently read the paper Attention selectively gates afferent signal transmission to area V4 and now wants to know if results can be reproduces. To figure out, data was provided (The data was generated artificially) from the same experiment and has kindly shared it with you. Now, it’s your duty as a good classmate to help answer the question.

Data set

All data required for this exercise can be found in: https://nc.uni-bremen.de/index.php/s/os5BC2nZeJGaJ8z

The main data set FlickerV1V4_LFP1kHz.mat comprises three arrays containing the data with a time resolution $dt = 1ms$ from $M=100$ trial repetitions over $T$ time bins:

flicker_signals (M x T x N) – flicker time series for stimuli $A$ and $B$

V1_lfp (M x T x N) – LFPs for $N=100$ V1 electrodes

V4_lfp (M x T) – LFP for one V4 electrode

Mandatory Tasks

1.

The goal of this project is for you to practice the process of scientific research. Before getting into the tasks, try to think thoroughly about the data set you have and the experiment by which this dataset is obtained. What questions can you investigate having such data? What methods can you propose to utilize in order to explore these questions? How can you handle your data and write your code for having a well-structured solution with re-usable components, which is easy to understand for outsiders (which include YOU after not having looked into the code for two weeks)?

We will ask some of you to present your thoughts on this task before you are going to implement the data analysis.

2.

Some scientists believe that (selective) information processing in the visual system is strongly linked to oscillatory and coherent neural activity, which is modulated by attention. Spectral analysis helps you in discovering and pinpointing oscillatory components in neural signals.

Use the wavelet transform (see A Practical Guide to Wavelet Analysis) to compute the trial-averaged amplitude spectrum over the whole trial duration and plot it. Make sure to transform your trials to have zero mean and unit standard deviation before applying the wavelet transform. Consider the parameters to choose for the analysis – what would be an appropriate frequency range and spacing (linear or logarithmic)?

Which part of your result can you trust? – include the cone-of-influence into your plot. Does the presentation of a stimulus generate oscillatory activity, and if yes, in which frequency band?

3.

{: .topic-optional} Don’t normalize the time series!

If appearance of stimuli is accompanied by enhanced oscillatory activity, you can use this dependence to determine regions of interest (ROIs) in the V1 neurons, i.e. sites that have the two visual stimuli in their receptive fields.

To do so, you can calculate the power spectrum in a specific frequency band for each channel and compare the oscillatory power across the neural population. Plot your results in a $10\times10$ grid. Where are the regions of interest? Select two neurons exhibiting the most significant power and report their corresponding indices.

4.

After performing the previous task, you might know which neurons are driven by visual stimuli, but not which site is responding to which individual stimulus [ie, flicker A or flicker B].

Find out by computing phase coherence and spectral coherence between the two selected neurons in V1 with each flicker signal. Compare the results.

For two wavelet-transformed signals with $a_1(t,f)$, $a_2(t,f)$ being the complex wavelet coefficients, the phase coherence $p(f) \in [0,1]$ is given by:

Similarly, you can compute the spectral coherence of these signals. The spectral coherence $c(f) \in [0,1]$ is given by:

$T$ contains time and trials.

5.

In the experiment, attention was devoted to one of the visual stimuli. You do not know to which one, but you know that V4 will selectively respond to the attended stimulus.

Compute the coherence of the V4 signal with each of the flicker signals. We first start with the phase coherence, and next apply the spectral coherence. Which of the two flicker stimuli is attended?

6.

{: .topic-optional} Don’t normalize the time series!

You might have observed that also V1 activity is modulated by attention (explain which result of your previous analysis supports such a statement!). How well can location of attention be decoded from one recorded electrode?

Here you will use some machine-learning techniques to classify attended against non-attended signals based on V1 LFPs. For this purpose, you have been provided with:

V1_LFP1kHz_NotAtt_train.npy and V1_LFP1kHz_Att_train.npy

which contain training data sets for 100 trials recorded from one specific channel, corresponding to the non-attended and attended conditions, respectively. After applying any classifier, you can use V1_LFP1kHz_NotAtt_test.npy and V1_LFP1kHz_Att_test.npy files to evaluate your results.

To get started, similar to what was previously done, load these signals and compute their time-frequency spectrum for different frequency bands. Since you need these spectra for all the following sub-tasks, it is recommended that you save your results in separate files.

7.

{: .topic-optional} Don’t normalize the time series!

Implement the ROC analysis as your own function/module. Then compute the ROC curve for different frequency bands. Plot the ROC accuracy over different frequencies: in which band do you get a better performance?

Voluntary Tasks: Preprocessing

For the main tasks, we provided you with data that was already properly preprocessed. Normally, several preprocessing steps ensure that your signals are clean and filtered in your frequency range of interest. Performing these steps can be trained on data we made ‘dirty’ for you:

1.

The data set V1_LFP1kHz_dirty.npy contains the LFPs which are still contaminated with a common source recorded on all channels (e.g., line noise). Load the data and remove the common signal using the singular value decomposition (SVD), then compare your cleaned signal with the LFPs in FlickerV1V4_LFP1kHz.mat.

2.

The data set V1_Raw32kHz.npy contains the raw signals sampled with a frequency of 32 kHz. These contain both low-frequency components (e.g. LFPs) as well as high-frequency components (e.g. spikes). Process these data by applying a lowpass filter to remove high-frequency spikes while preserving the underlying low-frequency neural activity. Optionally, you may downsample the signals to reduce computational load on subsequent analysis steps without sacrificing significant information.

Plot the filtered LFP signals to visualize the neural activity dynamics over time. Compare your results with the original signal. Be sure to label your axes appropriately for a clear interpretation of the results.

3.

Compute the firing rate of each neuron by analyzing the high-pass filtered raw signals from the V1 area. You can utilize appropriate techniques (such as any spike detection algorithms) to identify neuronal firing events. Once the firing rates are computed, plot a histogram of neuron firing distribution to visualize the variability across the neural population.

Concerning the memory of your computer

The shape of V1_Raw32kHz.npy is (trials=100, time=48000, channels=100) and the file is 3.6 GB in size.

Depending on the memory of your computer you can do this:

import numpy as np
data = np.load("V1_Raw32kHz.npy")
print(data.shape)

Or if your computer says no, you can load each trial individually like this

import numpy as np
trial_id: int = 0
data = np.load("V1_Raw32kHz.npy", mmap_mode="r")[trial_id, ...][
    np.newaxis, ...
].copy()
print(data.shape)

Voluntary Tasks: More Information!

Combining information contained in different frequency bands or recorded from different electrodes potentially improves classification accuracy. Here we use different methods to strive for such an improvement, and use statistical tests to determine whether results are (really) significantly different.

1.

{: .topic-optional} Don’t normalize the time series!

Based on your results from ROC analysis, select the best frequency bands for classification and use this range for the following tasks. Classify your data with the K-nearest neighbors (k-NN) and support vector machine (SVM) algorithms in the selected frequency bands. You can adjust the hyperparameters of these algorithms as needed to optimize classification performance. What is the best accuracy that you can achieve? Which method performs better?

2.

Now, you have applied three different techniques to classify your data (namely ROC, k-NN, and SVM). These techniques have different performances, but is this difference statistically significant? Apply the Fisher exact test for these methods against each other. Is there a method significantly better than the other one?

The source code is Open Source and can be found on GitHub.