Classification and prediction based on EEG signal have been accessed and evaluated in many research contexts during the past decades; however, there is currently no established way to classify a user's biological state with a high degree of accuracy in real-time. Applying cutting-edge technology, such as machine learning and deep learning algorithms to EEG signal turns out to be a new trend in general EEG classification offline. In our project, we will utilize the ML Engine and Cloud Storage provided by Google Cloud Platform and develop the possibility of a real-time EEG emotion classifying system for use in subversive real-world mobile application for marketing research.
What it does
Our project classifies a user's emotional state after only a short training time and with a sufficiently low inference time to be used for a biofeedback system. Marketing professionals can use this system to measure emotional state as their target audience experiences advertisements in the real-world (a concert, train-station, etc.)
How we built it
We set up the project on Google Cloud Platform and enable ML computing API with a Jupyter notebook. Then we set up google storage and upload data to bucket. The data came from the open DEAP dataset. Next we grant the read access of the bucket to GCP project so that those two Google products can connect together. The data came from the open DEAP dataset.
Challenges we ran into
We first tried to implement a deep learning architecture that was unable to perform better than random binary classification. This is because the algorithm needs to train on the individual. (And indeed no prior group has made a classifier that generalizes to anyone’s brain.)
Accomplishments that we're proud of
Our training time is short enough not to worry about training being a limiting factor, and the only preprocessing we used can be implemented in silico instantly. Therefore, the algorithm is performing near-instantaneous inference as well.
What we learned
We have to put signal through a transform to a power spectrum before end-to-end learning can be applied. If not, the signal does not correlate well. Thank you to the object extraction help session for this idea!
What's next for HappyEEG: Biofeedback for Strategic Marketing
Such a system could have implications in medicine, but we would like to develop it into a real-time bio-feedback system for the development of engaging advertising. Currently, advertising professionals engage volunteers to collect subjective feedback about a proposed advertisement. Our system could pick up on much more subtle feelings and desires from a targeted audience--including even subconscious desires such as jealousy of a luxury good. Furthermore, because we stressed the ability of low inference time, we could even automate the development of advertisements by development a biofeedback reward function that maximizes an audience's engagement.