Research Task 2, Part 2 (Isabelle Greenberg)

Brain-Activity-Driven Real-Time Music Emotive Control:
Abstract:

• Current active music systems allow a user to control aspects like the playback and volume.
• This project will use the Emotiv EPOC headset to receive EEG data and map it to emotional states. Then, the music will transform based on the user's mood

Introduction:

• This description is very similar to the abstract

Background:

• "GoTo" classifies music systems by their ability to control playback, touchup, retrieval, and browsing
• This system will have two parts; a real time system and a system able to adapt based on the real time data

2.1 Emotion Detection

• Methods of emotion detection are voice and expression, skin conductance, heart rate, and pupil dilation
• By measuring the alpha and beta activity on the prefrontal lobe, indicators for arousal and valence can be measured
• This can classify emotions such as happiness, anger, sadness, and calm

2.2 Active Music Listening

• Interactive systems allow the listener to control the music like a conductor using hand gestures
• The parameters are set from high to low, and they can effect the music and the influence of the gestures being made

2.3 Expressive Music Performance

• The KTH music system has a set of thirty rules that control different aspects of expressive music
• The magnitude is controlled by parameter "k", and different combinations of k parameters can create different styles, stylistic conventions or emotional intention
• An example of this situation was shown in 2006 with an arousal/valence plane and 7 of the rules

3 Methodology

• First, the process would begin with the Emotiv Epoch headset getting data in the prefrontal cortex (the sensors are F3, F4, AF3, AF4)
• That cortex regulates emotion and deals with our experience when we are conscious
• Affective states in emotions result from activation or deactivation (this is arousals) and pleasure or displeasure (this is valence)
• The results are happiness, anger, relaxation, and sadness and each has a different combination of arousal and valence

3.1 Signal Reprocessing

• Alpha waves are dominant in relaxed and awake states of mind
• Beta waves are indicators of excited mind states
• The signal has to first be filtered to get these waves separated
• The sample waves are put through a logarithmic equation that I do not presently completely understand but will look into further. It finds the band power.

3.2 Arousal and Valence Calculation

• The arousals are found by adding the beta band power of F3 and F4 and dividing it by the sum of the alpha band power of the same electrodes. The valence is found with a different equation

3.4 Synthesis

• The use of the rules comes in after the pDM program is used to analyze the values found so far

3.4 Experiments

• Two types of experiments were performed, one while sitting and not moving and one while playing an instrument
• The subject was asked to try to change their emotion and then the results from the software would be checked to see whether it was correct

4 Results

• The EEG signal came through and the arousals were mapped out on a graph, but the signal was not very smooth

5 Conclusions

• The opportunities for EEG data to control music is possible, because the arousals and valence values do change in accordance to emotion
• However, the signal is not smooth and more experiments would have to be done for emotions other than happy or sad

Algorithm Notes:

Decision Trees:

• Different patterns are divided into single pattern subsamples in a decision tree
• When there is only one pattern, the decision tree has a leaf
• Threshold: Purity measure of each node to improve feature selection
• The gain is the information value of all the features minus the information value of the nodes nodes
• sample sets are divided into two sets and the threshold can affect how the tree functions
• The entropy is low, or technically zero, when there is all elements of the same class. This means a higher gain
• At present, this math is confusing. The logs are understandable but how the equation is set up is unclear
• The leaf does not need to be processed any further

Self Organizing Maps:

• The inputs are in a table separated by lines and governed by different maximum and minimums 
• The values are updated on a graph for all the input data and a longer process is gone through
• The neurons are displaced?

K-Means Algorithms:

• Proper to compact clusters, sensitive to outliers and noise, and only works with numerical attributes
• The iteration is repeated again and again by increasing a counter each time it runs
• If the classification of one scenario is less than the convergence, it will repeat from an earlier step
• The K is given a  value and it will find the same amount of clusters as in that value
• Uses the random defined means to find all the elements based on the elements nearby to them
• The different means shift around the elements, but it still classifies the elements

Artificial Neural Network Algorithm:

• The inputs are 1 or 0
• The middle is hidden and two neurons are the first hidden layer and another neuron composes the second
• Propagate is to find the value
• The step function in the neuron is a simple graph measured by an equation
• Same inputs can have different weights

Support Vector Machine Algorithm:

• Research hyperplanes
• The larger margin for a hyperplane makes that hyperplane better
• research weight vectors

Random Forest Algorithm

• The combination of learning models increases the classification accuracy
• This idea is behind "bagging" which is to average noisy and unbiased models to create a model with low variance
• Random Forest: Large collection of decorrelated decision trees
• Research "matrix"
• this involves many features
• random subsets using the different matrix to create class prediction

Hopefully more basic research will allow me to process these videos better.