A Virtual Internship during the lockdown, 2020.

The only way to do great work is to love what you do. If you haven’t found it yet, keep looking. Don’t settle. As with all matters of the heart, you’ll know when you find it.

Steve Jobs

Before summer I was worried about getting an internship anywhere due to the pandemic. No wonder, most of the companies canceled their on-site internship opportunities before April 2020. Since doing an Internship was compulsory according to UGC, I hurriedly opened my laptop and looked for companies offering virtual internships. Leading India Ai is an initiative by Bennett University that provides internships on-site. Luckily, due to the lockdown, they were offering remote internships as well. Immediately, I contacted the respective person who is responsible for the enrollment of students for this internship. There were two slots available, one starting in May and the other in July. I wanted to get things done as soon as possible so I chose the May slot.
If you’re thinking that I enrolled myself in this internship just to complete the requirement of compulsory internship, you’re wrong.
The website mentioned, “During the 6 weeks internship, you will be actively working on a Deep Learning based project.” I joined this because I found myself loving this subject “Deep Learning” from my previous semester. Yes, this was a summer internship on Artificial Intelligence and Deep Learning.

Day 1 to Day 3
A crash course was hosted to cover the basic concepts related to Deep Learning. This lasted from morning till evening, it was kind of hard to sit in front of the laptop for so long but yes it was informative. This being held, even a beginner could start off with DL projects.

Day 3 to Day 40
There were more than 100 teams consisting minimum of 3 interns each. We were a team of 5 members. Everyone from a different state, no two from the same. This helped me converse with different people from different states. It was fun and stressful as well. Everything has its pros and cons right?
So the project topic assigned to us was “Feature Extraction Techniques For Classification Of EEG Signals.” Looking at the topic, I was startled and feared a bit if I could complete the desired objective. The voices in my head, “I have never heard of EEG Signals before. Oh, I did, but I do not know how they work. What am I gonna do? Classification is fine, what about feature extraction? Which features to take into consideration?”

Oh, I was joking about the voices in my head, I was blank. Period.

Okay, jokes aside!

Back to the topic, Every team was mentored by a researcher at Bennett University. Researchers were assigned according to the topic, obviously. So the mentor provided us with some research papers, articles, and datasets related to EEG Signals.

So what are EEG Signals? it’s an efficient modality that helps to acquire brain signals correspond to various states from the scalp surface area. In easy words, EEG is a recording of the electrical activity of the brain from the scalp.

Signal Intensity: EEG activity is quite small, measured in microvolts.
The main frequencies of the human EEG waves are Delta, Theta, Alpha, Beta, and Gamma.

EEG is most often used to diagnose medical problems like epilepsy, sleep disorders, depth of anesthesia, coma, memory problems, etc. Further classification of these EEG signals, we can determine which state a person is in (happy, sad, disgusted, angry, etc.)

For this, we need to extract features from the EEG signals. A feature is nothing but a distinguishing property obtained from a section of a pattern. Extracting these features leads to minimizing the loss of important information from the given signal.

These are the steps to be followed to determine the emotion of a person.

First, Acquire the EEG signals using a device ( Emotiv EPOC+ )
Next, Removing the noise from the data. This is to remove unnecessary details in the dataset which causes a decrease in accuracy. Then the relevant features need to be extracted which determine the state of a person.
The project was divided into two parts: feature extraction, and classification. Three of us were given the feature extraction part and the other two were to do the classification part. We used DEAP dataset which consists of 32 channels. Out of which we used only 14 channels ( taking EMOTIV EPOC+ in mind ) to extract the features.

There are a variety of methods used to extract the feature from EEG signals. Among these methods are Fast Fourier Transform (FFT), Wavelet Transform(WT), Time-Frequency Distribution (TFD), EigenVector methods(EM), Auto-Regressive methods (ARM), and so on. Among these methods, we have used the Fast Fourier Transform(FFT) and Discrete Wavelet Transform (DWT).

FEATURE EXTRACTION TECHNIQUES

Fast Fourier Transform (FFT)

This is one of the techniques that employ mathematical tools to analyze EEG data. The characteristics of the EEG signal are computed with the help of power spectral density (PSD) estimation to represent the sample EEG sample signal. The characteristics waveforms of the EEG spectrum are contained in four frequency bands. PSD can be calculated using the Fourier transforming the estimated autocorrelation sequence that is found by nonparametric methods.

The accuracy of the FFT technique is 84%

By the end of May, my teammates and I were able to complete the desired objective of extracting features from the raw EEG signal dataset provided to us by our mentor.

Keep doing the hard things to be the best version of yourself. I feared because I knew nothing about EEG Signals but now I’m confident about it.

Complete the work you started, no matter how hard things get.
So this way, I completed my internship on June 19th, 2020.

BIBLIOGRAPHY
https://github.com/tongdaxu/EEG_Emotion_Classifier_DEAP