# 猴子的EEG数据

**Repository Path**: iDMatlab/monkey-eeg-data

## Basic Information

- **Project Name**: 猴子的EEG数据
- **Description**: 本数据为斯坦福大学的参赛数据集，数据来源于马斯克的侵入式的脑电数据，利用现有数据预测猴子的行为
- **Primary Language**: Unknown
- **License**: Not specified
- **Default Branch**: master
- **Homepage**: None
- **GVP Project**: No

## Statistics

- **Stars**: 0
- **Forks**: 0
- **Created**: 2022-03-03
- **Last Updated**: 2022-05-31

## Categories & Tags

**Categories**: Uncategorized

**Tags**: None

## README

 ** The neural data have been generously provided by the laboratory of Prof. Krishna Shenoy at Stanford University. The data are to be used exclusively for educational purposes in this course.** 

# BMI Spring 2022 Neural Decoding Competition Rulebook

## 1. Description of the Competition

In the coming weeks you will design a neural decoder designed to drive a hypothetical prosthetic
device. This is a realistic and difficult task in brain-machine interfacing. The task will take the form of
a competition to make it even more enjoyable and to encourage creative solutions.
You will be provided with spike trains recorded from a monkey's brain, as it repeatedly performs an
arm movement task. Your algorithm will estimate, from these data, the precise trajectory of the
monkey's hand as it reaches for the target.
This is a continuous estimation task. Over the course of each spike train, you must estimate the X & Y
position of the monkey's hand at each moment in time (while the training data includes the Z position,
you are not asked to estimate it). Caveats: “time travel” is banned – i.e. your decoder must be causal,
it may not use information from the future to estimate the position of the hand at the present. When
your algorithm is tested, it will be given data in chunks of increasing length in order to enforce this
(see 5. Implementation).
## 2. Competition Scoring & Prizes
There will be a small prize in this contest, which the team with the lowest error will win. The prizes
are highly desirable though as-yet unspecified.
The following error metric will be used to evaluate the performance of the submissions: your error
will be the root mean squared error (RMSe) of the trajectories you compute. The mean will be
taken across both dimensions, over all trajectories. The units of this error are cm.
If you don’t submit a code by the end of the 3rd lab session, your final RMSe score will be increased
by 10% on the leader board for the Competition.
## 3. Teams and Deadlines
### • Teams
This competition will be a team effort. Each team will consist of 3-4 students (we will accept
exceptionally groups of 5 if well justified). You are encouraged to form the teams on your own, but
we are happy to provide help in forming the teams. Any serious grievances or disputes regarding
teams should be addressed to the GTAs or lecturers as early as possible.
### • Report Guidelines
Each team will submit only ONE final report. A template is available on Blackboard. The file will
have ALL of the team members' names written at the very TOP of the file. Group reports due on the
day of the BMI exam. They need to be uploaded on Turnitin. BlackBoard will not accept submission
after this time. Feedback to be provided within two weeks.
Structure the report as described in the attached Template: maximum 4 sides of two- column A4 paper.
This includes references. All team members should co-author the report: you should meet to plan how
you will coordinate the writing. We encourage groups to use the free online collaborative Latex editing
tool Overleaf, which includes the template we are using. To keep your report within 4 pages, you will
need to write concisely: this is good practice. Do not go over length – 10% will be deducted from your
marks if you do so. Also, please stick to the recommended font and font size. A 10% penalty mark will
be taken off the report if the team doesn’t participate to the competition.The neural data have been generously provided by the laboratory of Prof. Krishna Shenoy at Stanford University. The
data are to be used exclusively for educational purposes in this course.
The contributions of each member should be concisely noted in a section at the end. If team members
do not contribute, note this as well. Order your authors alphabetically in the author list. Team members
will normally receive the same, joint, mark for the report, however we reserve the right to bump
individual members marks up or down. All members should upload a copy of the group report to
Blackboard: the feedback will be marked up on the first author of the report (author with surname
closest to ‘A’): that person should share it with the other members of the group.
Describe the approach you took to solving the problem. Describe your methods and results, including
figures and/or tables describing the performance of your algorithm. In the Discussion, critically discuss
the performance of your algorithm and how it could be improved. How does it compare against other
approaches? A First-Class report would be, essentially, of a standard acceptable for submission to a
conference such as the International IEEE EMBS Conference on Neural Engineering.
### • Algorithm submission process
§ During the lab sessions, please submit your training function to the GTAs along with
your position estimator function (i.e. two separate functions) as two separate files,
preferably in a zip file. Please name the .zip file with your team name.
§ You will need to make sure your functions work. You can test this by running it
through the test functions provided on Blackboard. The format required is further
described in the comments of the template functions.
§ Your algorithm must take no longer than 5 minutes to run, as tested.
### • Deadlines
§ Homework (see A. Pre-Comp Homework). – To finish by the end of the first session,
22nd of February 2020.
§ SUBMISSION: Algorithm – 11 am GMT, Thursday 22nd March 2020.
§ SUBMISSION: Group report – noon on the day of the BMI exam.
## 4. Data
### • Overall Content
The data contains spike trains recorded from 98 neural units while the monkey reached 182 times along each of
8 different reaching angles, as well as the monkey’s arm trajectory on each trial. On each trial, both the neural
data and arm trajectory are taken from 300 ms before movement onset until 100 ms after movement end. The
trials of different reaching angles were interleaved; and the neural data includes both well-isolated single-neuron
units (~ 30% of all units), as well as multi-neuron units. The set has been divided into a training dataset of 100
trials per reaching angle and a test file of 82*8 trials. The former can be found on Blackboard as a .mat file,
while the latter is not accessible by students.
### • .mat File
The .mat file has a single variable named trial, which is an array of dimensions 100(trials) x 8(reaching angles).
The trial array contains 100*8 struct data structures which contain 3 fields: trialId, spikes, handPos. The first
is the trial ID number and is therefore a of little consequence. The latter two are arrays containing the neural
data and the hand position traces respectively.
### • Neural Data
The spike train data takes the form of a time-discretized sequence of zeros and ones with time steps of 1 ms.
Each value in the array therefore takes the binary form 0/1 and represents the presence or absence of a spike in
its neural unit in the 1 ms time window, or bin. A zero indicates that the unit did not spike in the 1 ms bin,
whereas a one indicates that the unit spiked once in the 1 ms bin. Thus, a spike train of duration T ms is
represented by a 1*T vector. The spike train recorded from the ith unit on the nth trial of the kth reaching angle isThe neural data have been generously provided by the laboratory of Prof. Krishna Shenoy at Stanford University. The
data are to be used exclusively for educational purposes in this course.
accessed as: trial(n,k).spikes(i,:) where i = 1, . . . , 98, n = 1, . . . , 100, and k = 1, . . . , 8.
### • Trajectory Data
In this task, the monkey reached to targets on a fronto-parallel screen. Most of the arm movement was in the
plane of the screen along the horizontal directions: handPos(1,:) and handPos(2,:). The movement perpendicular
to the plane of the screen, handPos(3,:), was relatively small.
The three-dimensional arm trajectory recorded on the nth trial of the kth reaching angle is contained in
trial(n,k).handPos, which is a 3*T matrix of the hand position (in mm) at each 1 ms time step. On each trial, the
data in spikes and handPos are aligned in time.
The indices k = 1, . . ., 8 correspond to the reaching angles (30/180π, 70/180π, 110/180π, 150/180π, 190/180π,
230/180π, 310/180π, 350/180π) respectively. The reaching angles are not evenly spaced around the circle due
to experimental constraints that are beyond the scope of this problem set.
## 5. Implementation
The competition will be run in Matlab. While you are welcome to use other programming languages
for your own edification, non-Matlab submissions will not be graded.
Your submission for each task will take the form of a MATLAB function. The function will accept
two arguments: a set of training data (that includes reaching angles and trajectory data), and a set of
test data (which only has spike trains and are not included in your training data). Your function will
then return the appropriate angle classification and position estimation.
You will be given a template (accessible through Blackboard), including a function declaration with a
description of the arguments it must accept and the data structure it must return. You may change the
variable names but not the function name nor the order of the arguments.
The model must be causal. To enforce this, the model will be scored at regular intervals of 20 ms: to
start, the first 320ms are fed in, then the first 340 ms, then the first 360, 380, 400, 420… until the whole
recording is eventually fed in to your model. Reading the code and comments in the template will
clarify the implementation significantly. The test script run by the GTAs has the same format and feeds
data into your algorithm in the same way but uses the test data to evaluate your model.
## 6. Inviolable Laws
• You MUST cite every resource you use.
• Your submissions MUST implement the interfaces in the Matlab templates exactly as
specified.
• You MUST put the names of ALL team members on the top of EVERY file you submit.
• While you may use existing (cited) algorithms, all submitted Matlab code MUST be
entirely your own. Exceptions are the Matlab template files, any code posted by the
lecturers or GTAs on Blackboard, and the calling of standard Matlab functions. You cannot
use any MATLAB toolbox (such as the signal processing toolbox) to win the competition.
All submitted code must be written by you and all implementations of algorithms you've
learned about or researched must be written by you. You can use imported/downloaded
toolboxes and libraries to explore different options and methodologies, generate interesting
results for discussion in your report, and have fun – BUT YOU CAN'T SUBMIT CODE
THAT RUNS USING FANCY TOOLBOXES. Late submissions will be ignored. Your
submission may be updated online, so it is suggested that you submit early and then
continue updating your submission until the deadline if you so desire.The neural data have been generously provided by the laboratory of Prof. Krishna Shenoy at Stanford University. The
data are to be used exclusively for educational purposes in this course.
## 7. Final bit of Advice
Feel free to use whatever algorithms you like, as long as you write the code yourself. In particular, you
are encouraged to attempt to design your own. You have also been exposed to several algorithmic
ideas throughout the lectures, and you may mime research papers, professors, textbooks, etc. The sky
is the limit. All code submitted must be your own and plagiarism checks are implemented.
Be warned, however, that more mathematical artillery doesn't necessarily result in a better solution.
Consider implementing a simple idea early on and then taking time to improve on it, rather than using
up all your time designing a complicated but untested idea. Remember also that the ideas described in
research papers rarely work exactly as specified.
A. Pre-Competition Homework
Before you start with the BMI competition you will try to get comfortable with the dataset by playing
around with it. You will visualize parts of the dataset by writing and performing calculations in Matlab.
First, download the monkeydata_training.mat file from \Course Content\The Competition location in
Blackboard. Start Matlab and set the working directory to the location of the monkeydata_training.mat
file. Load the data set by either dragging the file into the Matlab workspace or by using the following
command:
 >> load('monkeydata_training.mat');_ 
Read the section 4 ‘Data’ from this document to familiarize yourself with the fields in your
workspace.
Once you are comfortable with the data fields in your workspace, try the following exercises:
#### 1) Familiarize yourself with raster plots and for a single trial, compute and display a population
raster plot. A population raster would have time (bins) on the x-axis and neural units on the yaxis. A sample raster plot can be found here:
http://en.wikipedia.org/wiki/File:Sample_raster_plot_from_Brian_neural_network_simulator.jpg
2) Compute and display a raster plot for one neural unit over many trials (suggestion: try using
different colours for different trials to help with the visualization).
#### 3) Familiarize yourself with what peri-stimulus time histograms (PSTHs) are and compute these
for different neural units. Refer to http://icwww.epfl.ch/~gerstner/SPNM/node7.html for more
information on PSTHs. Try smoothing your histogram to get a continuous rate estimate.
#### 4) Plot hand positions for different trials. Try to compare and make sense of cross-trial results.
#### 5) For several neurons, plot tuning curves for movement direction. Tuning curves measure
directional preference of individual neural units. Different neural units will have different
directional preferences.
Get the tuning curve, by plotting the firing rate averaged across time and trials as a function of
movement direction. To get an idea of the variability, compute the standard deviation across
trials of the time-averaged firing rates. You can use the Matlab command ‘errorbar’ to plot the
tuning curve with error bars indicating the variability.
### 6) Think about other ways that you can play around and visualize the data in order to understand
the nature of the experiments.The neural data have been generously provided by the laboratory of Prof. Krishna Shenoy at Stanford University. The
data are to be used exclusively for educational purposes in this course.
#### 7) If you are feeling clever: implement the population vector algorithm and use it to predict arm
movements.