170 likes | 301 Views
Assignment 2: reverse correlation. Outline. The assignment requires you to Write code to produce graphs Make observations from the graphs Draw conclusions. Coding. Coding is in MATLAB. I will provide you with templates that provide you with: A list of MATLAB functions to use
E N D
Outline • The assignment requires you to • Write code to produce graphs • Make observations from the graphs • Draw conclusions
Coding • Coding is in MATLAB. • I will provide you with templates that provide you with: • A list of MATLAB functions to use • Comments describing the flow of operations
Some Coding Tips • It is important that you know how to use the debugger. • Use the MATLAB Help facility. • You should generally never have a loop (or nested loop) that involves more than a few hundred iterations.
Dataset • We will be using a portion of the Neural Prediction Challenge Dataset • Responses of V1 neurons to natural vision movies in awake behaving macaque • Both neural responses and visual stimuliare provided • Available at http://neuralprediction.berkeley.edu/ • But you can download the files you need from the course website. We will be analyzing a particular neuron (R0221B)
Submission Details • You will submit a short lab report on your experiments. • For each experiment, the report will include: • The code you developed • The graphs you produced • The observations you made • The conclusions you drew
Graphs • The graphs you produce should be as similar as possible to mine. • Make sure everything is intelligible!
Due Date • The report is due Wed Mar 23
Reverse Correlation • Raw stimulus response cross-correlation: • Now represent the kernel h as an mxTmatrix, where • Correction for temporal stimulus bias: • Correction for spatial stimulus bias: • But this doesn’t work, because there are too many coefficients in Qss to estimate, and too little power in the high frequencies of the stimulus to estimate them.
Solution: Regularized Inverse • Use SVD decomposition: • Where U and V are orthonormal rotation matrices and S is a diagonal scaling matrix carrying the eigenvalues of Qss • The eigenvalues represent the power of the autocorrelation in each of the underlying principle directions (eigenvectors).
Regularized Inverse • Once the SVD decomposition is computed, taking the inverse is easy. • However, this inverse is unreliable, because noisy eigenvalues in S near 0 result in large noisy values in S-1. • To avoid this, only take the largest eigenvalues from S, and set the remaining diagonal elements of S-1 to 0.
Firing Rates Histogram KDE