Computational Neuroscience Homework 3 Solution

Instructions:




Prepare a report (including your answers/plots) to be uploaded on Moodle.



The report should be typeset (for lengthy derivations, the solution can be scanned and embedded into the report).



Show all the steps of your work clearly.



Unclear presentation of results will be penalized heavily.



No partial credits to unjusti ed answers.



Use Matlab or Python for computations.



Return all Matlab/Python code that you wrote in a single le.



Code should be commented, code for di erent HW questions should be clearly sepa-rated.



The code le should NOT return an error during runtime.



If the code returns an error at any point, the remaining part of your code will not be evaluated (i.e., 0 points).





















Question Points Your Score







Q1 20







Q2 30







Q3 30







Q4 20







TOTAL 100




Question 1. [20 points]




Responses of a V1 neuron in an awake monkey are provided in the le hw3_data1.mat. This le contains two variables resp1 and resp2 that represent the response levels during a passive xation condition and those during a memory task, respectively. Answer the questions below.







Using ordinary least-squares (OLS) method, t a linear model y = a x + b, where the independent variable (i.e., regressor) is resp1 and the dependent variable (i.e., output) is resp2. Show the original data on a scatter plot, and show the model t with a line on



the same graph. Report the estimated model parameters on the graph. Find the explained variance, unexplained variance, and the coe cient of determination (R2) for the linear model. Compare R with Pearson's correlation coe cient between resp1 and resp2.




Using OLS, t a linearized second-order model y = a x2 + b x + c, where resp1 is the regressor and resp2 is the output. Show the original data on a scatter plot, and show the model t with a curve on the same graph. Report the estimated model parameters



on the graph. Find the explained variance, unexplained variance, and the coe cient of determination (R2). Compare R with Spearman's correlation coe cient between resp1 and resp2.




Fit a parametric nonlinear model y = a xn +b, where resp1 is the regressor and resp2 is the output. Use lsqcurvefit function to nd the model parameters, starting the iterative



algorithm at two di erent initial values: fa; n; bg = f1; 1; 0g and f10; 7; 100g. Show the original data on a scatter plot, and show the two model ts with separate curves on the




same graph. Report the estimated model parameters on the graph. Find the explained variance, unexplained variance, and the coe cient of determination (R2) for both models. Which model performs better and why?




Fit a nonparametric nonlinear model using nearest-neighbor regression, where resp1 is the regressor and resp2 is the output. Show the original data on a scatter plot, and show



the model t on the same graph. Find the explained variance, unexplained variance, and the coe cient of determination (R2). Interpret the meaning of your R2 measurement




Question 2. [30 points]




A two-alternative forced choice (2AFC) experiment is conducted, in which a subject views two stimulus arrays of potentially di erent intensities side by side. In each trial, the subjects is assigned the task to determine which array contains the target stimulus (only one of the arrays contain the target in each trial). The probability of giving a correct answer on each trial (i.e., the psychometric function) is given by:




pc(I) =
1
+
1
(I; ; )
(1)








2
2



where (I; ; ) is the CDF of a Gaussian random variable with mean and standard de-viation , evaluated at stimulus intensity I. Answer the questions below.




a) Plot the pscyhometric functions for f ; g equal to f6; 3g and equal to f3; 4g, for I 2 [1 10]. Describe and interpret the di erences between the two functions. Is the range of pc(I) appropriate based on the description of the 2AFC experiment?




Write a function [C,E] = simpsych(mu,sigma,I,T) that takes two vectors I and T of



the same length, containing the stimulus intensities and number of trials for each intensity, respectively. The function should simulate random draws from pc(I). Outputs are a vector C that contains the number of trials correct out of T at each stimulus intensity I, and a




matrix E that contains the trial result at each stimulus intensity and each of T trials at that intensity. Show a scatter plot of C/T versus I, and a curve plot of pc(I) on the same graph, assuming T=ones(1,7)*100, I 2 [1 : 1 : 7], = 5 (mu), and = 1:5 (sigma).




Write a function nll = nloglik(pp,I,T,C) that returns the negative of the log-likelihood of parameters pp.mu, pp.sigma given an experimental dataset of I,T,C. Show a contour



plot (using 50 contours) of the negative log-likelihood of the dataset generated in part b (i.e., using the I,T,C vectors from the previous part), for all pairs f est; estg where




est 2 [2 : 0:1 : 8] and est 2 [0:5 : 0:1 : 4:5]. Determine the best tting parameters f est; estg by visual inspection.




Using the function fminsearch, nd a more precise estimate of f est; estg that min-imizes the function nloglik. Hints: You can call fminsearch with an initial starting



point of {2,2} for f est; estg. You are strongly recommended to use the fit.m package (the fminsearch interface coded by Geo Boynton) that is available on Moodle. Be careful when calling this function, since there is also a native fit.m function in Matlab.




Determine con dence intervals for the parameter estimates using the bootstrap tech-nique. For each stimulus intensity, generate bootstrap samples by resampling the 100 trials of that intensity in the original data (i.e., resample E). A set of bootstrap samples for all



stimulus intensities consititutes a resampled dataset; re t the model to this dataset using fminsearch. Perform 200 bootstrap iterations, and plot the histograms of f est; estg. Find the 95% con dence intervals.




Question 3. [30 points]




Blood-oxygen level dependent (BOLD) responses of a neural population in human visual cortex are provided in the le hw3_data2.mat. This le contains a variable Yn that represents 1000 response samples. There is another variable Xn that represent 100 regressors that may explain the responses. For all parts, the proportion of explained variance (R2) should be calculated as the square of Pearson's correlation coe cient between measured and predicted responses. Answer the questions below.







Use the ridge regression method to t regularized linear models to predict noisy BOLD



responses as a weighted sum of given regressors. Perform 10-fold cross-validation to tune the ridge parameter ( 2 [0 1012]) based on model performance. (Hint: Vary the ridge parameters logarithmically.) Note that for = 0, the model obtained with ridge regression is equivalent to the OLS solution. For each cross-validation fold, do a three-way split of the data: select a validation set of 100 contiguous samples, a testing set of 100 samples (that




immediately precede the validation set assuming circular symmetry), and a training set of length 800 samples. Fit a separate model for each using the training set. Find R2 of each model on the testing set. Separately estimate R2 of each model on the validation set. Plot the average R2 across cross-validation folds, measured on the testing set as a function of . Find the optimal ridge parameter opt that maximizes average R2. Find the model performance by calculating the average R2 across cross-validation folds, measured on the validation set for opt. Plot R2 curves obtained on testing and validation data for all values. Interpret your results.




Determine con dence intervals for parameters of the OLS model from part a (i.e., the model obtained for = 0). Generate bootstrap samples from the 1000 samples in the original data (resample both the regressors and the responses the same way). Perform 500 bootstrap iterations, and re t a separate model at each iteration. Plot the mean and 95% con dence intervals of the parameters in the same graph. Identify and label on your plots, the model regressors which have weights that are signi cantly di erent than 0 (at a signi cance level of p < 0:05).



Determine con dence intervals for parameters of the regularized linear model from part a (i.e., the model obtained for opt). Generate bootstrap samples from the 1000 samples in the



original data (resample both the regressors and the responses the same way). Perform 500 bootstrap iterations, and re t a separate model at each iteration using opt found in part a. Plot the mean and 95% con dence intervals of the parameters in the same graph. Identify and label on your plots, the model regressors which have weights that are signi cantly di erent than 0 (at a signi cance level of p < 0:05). Compare the results to those in part b.




Question 4. [20 points]




A series of neural response measurements are provided in the le hw3_data3.mat. Answer the questions below to examine the relationship between these measurements. Provide plots whenever possible.







Responses from two separate populations of neurons are stored in the variables pop1 and pop2. We would like to examine whether the mean responses of the two populations are signi cantly di erent. The rst population contains 7 neurons, whereas the second population contains 5 neurons. Using the bootstrap technique (10000 iterations), nd the two-tailed p-value for the null hypothesis that the two datasets follow the same distribution. (Hint: If the two datasets come from a common distribution, is there any need to separate them?)



BOLD responses recorded in two voxels in the human brain are stored in the variables vox1 and vox2. We would like to examine whether the voxel responses are similar to each other, by calculating their correlation. Using the bootstrap technique (10000 iterations), nd the mean and 95% con dence interval of the correlation. Find the percentile of the bootstrap distribution, corresponding to a correlation value of 0. (Hint: Should you resample vox1 and vox2 independently or identically?)



Note that estimation of con dence intervals and hypothesis testing are dual problems. For the dataset examined in part b, use bootstrapping (10000 iterations) to simulate the distribution of the null hypothesis that two voxel responses have zero correlation. Find the one-tailed p-value for the two voxel responses having zero or negative correlation. Compare this to the result in part b. (Hint: Resample the datasets to break apart the correlation between them.)



The average BOLD responses in a face-selective region of the human brain have been recorded in two separate experiments. The responses of this region to building images (1st experiment) and face images (2nd experiment) are stored in the variables building and face for 20 subjects. Assume that the same subject population was recruited in both experiments. Use bootstrapping (10000 iterations) to calculate the two-tailed p-value for the null hypothesis that there is no di erence between the building and face responses.



Repeat the exercise in part d, but this time assuming that the subject populations recruited for the two experiments are distinct. Use bootstrapping (10000 iterations) to calculate the two-tailed p-value for the null hypothesis that there is no di erence between the building and face responses.