OpenViBe paper
Section 1:
The OpenViBE platform is a way to
design BCI systems.
Section 2:
BioSig - software library for BCI
research.
BCI1200 - system for BCI research.
BCI++ - framework for designing
BCI systems.
Section 3:
OpenViBE is a free platform for
designing, testing and using BCI. Very versatile, can adapt to various needs.
OpenViBE allows non programmers to
design BCI without having to write code. Also can connect and work with VR
systems easily.
Section 4:
4 types of users:
The developer - uses OpenViBE
adding new pieces of software. OpenViBE has software development kit. Can
either create new/modify old. Programmer.
The application developer - uses
the software programming kit to create standalone applications. Ex: VR
applications that the BCI can interact with. Programmer.
The author - uses easy process to
create BCI system. Basic signal processing knowledge etc, but non-programmer.
The operator - tests out BCI,
knows how to run it.
Section 5:
A- acquisition of training data
B- offline training
C- Online BCI loop:
1- brain activity measurment
2- preprocessing (filtering
out/enhancing the specific signals)
3- feature extraction (extraction
of features that describe relevant info embedded in the signals, such as the
power of signals in specific frequency bands. These features are gathered in a
feature vector.)
4- classification (the feature
vector is fed into an algorithm called the classifier. This assigns a class to
each vector, an identifier of the brain signal recognized.)
5- translation into a command
6- feedback (so the user can see
if he did it correctly)
Section 6:
Tools provided in the system:
The acquisition server - provides
interface to various kinds of acquisition machines (EEG systems, for ex.).
The designer - helps author build
scenarios using a graphical interface. The author has access to a list of existing
modules and can drag and drop them in the scenario window.
The 2D visualization features -
specific boxes that include brain activity related visualizations. Ex: power
spectrum, time frequency map etc.
Existing and pre-configured
ready-to-use scenarios - proposed to assist author. Ex: a scenario that enables
the user to spell letters using only brain activity.
Section 7:
The platform is able to interact
with VR applications. It has a complete library to build 3D plugins, and offers
the functionalities to load and modify a 3D scene based on input data. The
platform also can be used to visualize brain activity in real time.
Section 8:
The box concept
The kernel
Plug-Ins
(Internals, how OpenViBE works)
Section 9:
Examples of implementation
explained: handball game, use-the-force application
Section 10:
Tests proved efficiency of the
system.
Section 11:
Available on windows
Includes a lot of boxes and stuff
it can do
Section 12:
OpenViBE website
NueroSky MindWave user guide
I read through it all and despite
all the fancy jargon (or maybe because of it) I couldn't successfully connect
and start using mindwave. So below is all necessary information regarding
mindwave: (I copied/pasted these instructions from an email Mr. Lin sent me,
they are really clear and helpful)
For pairing the Bluetooth device:
1. Left click the little "triangle" ("show hidden icons") at the bottom right area of the Taskbar. You will see the Bluetooth icon in the pop-up menu.
2. Right click the "Bluetooth icon", and select "Add a Device".
3. Now turn on the NeuroSky Mindwave headset, and you will see it blinking.
4. You can push the switch further (the blue light will blink faster) for a few seconds to pair with PC.
5. If you push too long, the light will turn red, and you will need to turn the headset off and on again. Try a few times, and it will work.
6. Once the PC found it, click the "Mindwave icon" and then click "Next".
7. After loading the driver and configuring the computer, there will be message to tell you "This device has been successfully added to this computer". Now click "Close" and you are done.
1. Left click the little "triangle" ("show hidden icons") at the bottom right area of the Taskbar. You will see the Bluetooth icon in the pop-up menu.
2. Right click the "Bluetooth icon", and select "Add a Device".
3. Now turn on the NeuroSky Mindwave headset, and you will see it blinking.
4. You can push the switch further (the blue light will blink faster) for a few seconds to pair with PC.
5. If you push too long, the light will turn red, and you will need to turn the headset off and on again. Try a few times, and it will work.
6. Once the PC found it, click the "Mindwave icon" and then click "Next".
7. After loading the driver and configuring the computer, there will be message to tell you "This device has been successfully added to this computer". Now click "Close" and you are done.
For running OpenViBE with the
headset:
1. Click "Start > All Programs > OpenViBE > openvibe acquisition server. A terminal window will pop up and running some scripts.
2. At the end of the scripts, a new OpenViBE Acquisition Server window will pop up.
3. Change "Driver" to "NeuroSky MindSet" by clicking and selecting from the drop-down menu.
4. Click "Connect" and you will see some scripts (with green "INFs" running in the terminal window.
5. Click "Play". In a few seconds, you will see some "Blue Bars" changing at the bottom. It means that the OpenViBE has start acquire your brainwaves. You are all set!
1. Click "Start > All Programs > OpenViBE > openvibe acquisition server. A terminal window will pop up and running some scripts.
2. At the end of the scripts, a new OpenViBE Acquisition Server window will pop up.
3. Change "Driver" to "NeuroSky MindSet" by clicking and selecting from the drop-down menu.
4. Click "Connect" and you will see some scripts (with green "INFs" running in the terminal window.
5. Click "Play". In a few seconds, you will see some "Blue Bars" changing at the bottom. It means that the OpenViBE has start acquire your brainwaves. You are all set!
For runnning OpenViBE Designer:
1. Click "Start > All Programs > OpenViBE > openvibe designer. Another terminal window will pop up and running some scripts.
2. At the end of the scripts, a new OpenViBE Designer window will pop up. Now you are ready to follow the video tutorial and learn the OpenViBE step-by-step!
NueroSky Mindwave and OpenViBE
video
·
Try
wearing the device backwards and compare results.
OpenViBE video tutorial
seems simple enough, but I tried
following the instructions several times and it won’t work. (for example, I
can’t seem to find the sample ghp files to run.)
Classification algorithms video series
Video 1:
Decision trees algorithm
Building a decision tree
When you have a set with samples
of different patterns, set a threshold and divide it, building a tree (kinda
like a family tree style). For ex: all elements higher than 90 go right, all
lower than 90 go left. Now you purposely want to build your threshold so as to
isolate different features. For ex: once you've divided your data in the first
threshold, you see that all data on the right is blue, while the data on the
left is red and blue. Success! You've established that all data higher than 90
is blue. Now the right leaf of your decision tree is done, and will not extend
any further. Now create a new threshold for the left side (ex: all data older
than a year right, all other left), and aim to create a leaf that will not
continue. Do so until your last node yields two leaves that stop.
·
Leaf:
branch on decision tree that does not continue (no entropy) (it's a 'pure'
node)
·
Node:
branch with entropy, that will continue to be divided in two so as to reduce
the entropy
·
Threshold:
a purity measure to improve the feature of selection.
·
Entropy:
how uncertain are we of the outcome?
H(P1, P2,...Pn) = -K £ Pi Log Pn
(-P1/sum x log P1/sum)+(-P2/sum x
log P2/sum) etc
·
Gain:
how much entropy was reduced by some test?
Info(all features) - info(nodes) =
gain(feature)
So basically said set is a bunch
of jumbled together elements, thus containing high entropy. We set a threshold;
we have now divided the set and created a pure leaf. We have lowered the
entropy. What is the exact gain of this threshold then? (How much entropy did
this specific threshold reduce?) calculate this by subtracting the entropy
value of the (node + leaf) from the entropy value of the data before the
threshold was set. Your answer will be the gain of the leaf you just created.
(That's because if you've in fact created a pure leaf, the entropy of the leaf
is zero. Thus, you're just subtracting the entropy of the node from the
previous entropy value. If you've created two nodes, though, you might still
have a gain, because the nodes might be better divided and filtered down.
That's why you do (node+leaf), cause in case you haven't actually created a
leaf it ends up being (node+node).)
Ex: follow along with his video,
it's pretty clear once you get it.
Video 2:
Self organizing maps
The inputs are data from a
specific set of neurons (let's say 2). They are on a graph.
1. Select a random input.
2. Compute winner neuron (which
neuron is closest to input).
3. Update neurons (the neuron will
be arranged closer to the input. But because both neurons are connected, in a
sense, both neurons will kind of be rearranged).
4. Repeat for all inputs
5. Classify input data (all inputs
closer to neuron 1 will be classified in the same groups it. All inputs closer
to neuron 2 will be classified in the same group as it.).
Video 3:
K-Means algorithm
·
Proper
for compact clusters
·
Sensitive
to outliers and noise
·
Uses
only numerical attributes
Input: set of feature vectors
K: number of clusters to be
detected
Convergence: number of input that
stays in the same cluster as it was last
1. Define an initial (random)
solution as vectors of means Mt = (M1, M2,...MK)
2. Classify all input data
according to initial solution (is it closest to M1, M2, etc etc)
3. Use classifications obtained in
step 2 to recompute vectors of means as M(t+1)
4. Update t=(t+1)
5. If ||Mt - M(t-1)|| <
convergence , use Mt as solution. If not go back to step 2.
·
(How
to know what is convergence threshold??)
Video 4:
Neural networks algorithm
Propagate the neural network: go
from input to output.
Follow along video example, it's
pretty simple.
Two inputs: 0.0 and 1.0. We will
plug both into the step function, and receive an output. We will plug both into
the logistics function, and receive an output. Now we will forget the two old
inputs and just take these two new outputs and they will become our new input
values. Now we continue to propagate it by plugging both into the linear
function, and receiving an output. This is the final output.
A) step function
1. Each input will receive a
weight. In this case, 0.0 has a weight of 0.3, and 1.0 has a weight of 0.7.
2. Multiply each input by its
weight. Add both results. In this case, (0.0x0.3=0.0)+(1.0x0.7=0.7)=0.7
3. Now plug your x value - 0.7 - into
the activation function for the step function. (Imagine plotting a value on a
graph for x and seeing what the y value is.)
4. In this case, f(x) = 1.0. This
is our output. It will later become our new input.
B) logistic function
1. Again, each input will receive
a new weight. In this case, 0.0 has a weight of 0.1 and 1.0 has a weight of
0.1.
2. Multiply each input by its
weight. Add both results. In this case, (0.0x0.1=0.0)+(1.0x0.1=0.1)=0.1
3. Same as before. X value is 0.1,
plug into activation function for logistic function.
4. F(x) = 0.2. Our output for this
function (and our eventual new input) is 0.2.
C) linear function (now we are
dealing with new inputs, the outputs from the past functions.)
1. Each input receives a weight.
1.0 has a weight of 0.3, and 0.2 has a weight of 0.5.
2. Same as before.
(1.0x0.3=0.3)+(0.2x0.5=0.1)=0.4
3. X value is 0.4. Plug into
activation function.
4. F(x) = 0.7. This is our final
output.
Video 5:
SVM algorithm
·
For
linearly separable binary sets
·
Hyperplane
(imagine a line)
The goal is to design a hyperplane
that classifies all training vectors into two classes.
Imagine two clusters on a graph.
Now we need to draw a line to show a division between the two clusters. Many
lines can be made. (let's say cluster one is above 20 and cluster 2 below 10.
Now we can make a line y=12 and it divides both, but we can also draw y=13 and
it divides it.) we want the line with the most margin between both classes. (In
this case, y=15 is the best bet.)
The hyperplane (line) has an
equation. If we plug in the values of the class above the hyperplane into the
hyperplane's equation, it will yield values greater or equal to 1. If we plug
in the class below the hyperplane, it will yield values less than or equal to
-1. (In this case, when we plug in any value from the cluster under 10 into the
(y=15) line's equation, it will give us a value less than or equal to -1, and
if we plug in any value from the cluster above 20, it will give us a value
greater than or equal to 1.) so the distance from the closest element until the
hyperplane is always at least 1.
We are trying to maximize the
separability and maximize the margin. Use equation.
Video 6:
kNN algorithm
Given N training vectors, kNN
algorithm identifies the k nearest neighbors of c, regardless of labels.
Ex:
·
K=3
·
Classes
a and o
·
Aim:
Find class for c.
c is another point on the graph.
We are going to determine if it falls into class a or o by finding its 3
nearest neighbors. We find that it's nearest neighbors are 3 elements, two of
class o and one of class a. Thus, c is classified as class o.
Remarks:
·
Choose
an odd k value for a two class problem
·
K
must not be a multiple of the number of classes
·
The
main drawback of kNN is the complexity for searching the nearest neighbors for
each sample (in cases with lots of elements).
When k=1:
Each training vector will define a
region in the feature space. Defines a voronoi partition. The distance between
any point in this space till the vector is always smaller than the distance of
the same point to a different vector (that's what makes it a voronoi
partition).
Video 7:
Random forest algorithm
·
The
combination of different learning models increases the classification accuracy.
·
Bagging:
to average noisy and unbiased models to create a model with low variance
·
This
algorithm works as a large collection of decorrelated decision trees.
How it works:
1. Assume matrix S of various
training samples. fA1---fAn, fB1---fBn, etc. (feature A of sample 1 till sample
n. Feature B of sample 1 till sample n.) last column specifies how many
features of the training class. (Ex: C1---Cn)
2. Create random subsets. (For ex:
all features of samples 12, 15, and 35 in one subset. All features of samples
2, 6, and 20 in another subset. Etc)
3. Create decision trees using
your random subsets. (So let's say you now have 4 trees you created using four
random subsets.)
4. Predict the classification for
each sample. (Let's say 2 trees are classified as class 1, one as class 2 and
one as class 3.)
5. Count the votes. (In this case,
the information will be classified as class 1.)
Video 8:
Supervised and unsupervised
classification algorithm
·
A
priori information: information that is justified by knowledge. We know
something about this information.
·
When
dealing with supervised classification, we are dealing with a priori information.
Ex: points on a graph. We are
trying to classify them. The information we know: three areas of training
patterns. Now we can classify all other points based on these areas of
patterns.
·
Examples
of supervised algorithms: decision trees, random forests, kNN, svm, neural
networks
·
When
dealing with unsupervised classifications we do not have a priori information
available.
To find groups of points we use
clustering. Ex: a cluster of points are close to each other so we group them together.
The question is always how many clusters to define. Sometimes it is given (for
example in k-means, the k stands for how many clusters we are supposed to
find.).
·
Examples
of unsupervised algorithms: k-means, self organizing maps, expectation-maximization,
isodata, hierarquical clustering
Video 9:
MLE algorithm
we are trying to find class
conditional density given a feature vector and given C sets.
Ex: c=3. We have three classes -
red, blue, green. Every class has several values. We are trying to find the
parameters of each class. So let's say for class red,
·
average
x value = -0.07.
·
The
standard deviation in x values = 3.02.
·
The
average y value = 4.83
·
The
standard deviation in y values = 0.98.
We do the same for class blue and
green. This defines the parameters of all classes.
So if we're trying to find the
maximum likelihood estimation of a random point, we use the class conditional
density equation. We plug in the x coordinate and get an answer, we plug in the
y coordinate and get an answer, and we add the two. We repeat with all classes.
The class that yields the highest result is the class that this point has the
maximum likelihood of being in.
