1 of 20

EE-JAMS

(Joint Affective Measurement System)

By CruX @ UCLA

Presentation Notes

Many technical issues with displaying

Need to decide:

Which displays we’re using (Zoom screen sharing?)
The order in which we’re doing the demos

So it seems like we’re going Emotion EEG demo -> slides -> cap headset

Should have shown the back up demo for the cap headset

Rehearse more

Need to have back ups for when demos break (so for example we should have the video of the cap headset working on the slides)

Questions judges asked: What are the features we’re using to classify the data?

Our answer: due to the black box nature of CNNs, it gets lost

This kind of translates to us saying we don’t really know
We should probably explain we’re feeding in 3-second chunks of raw EEG data into the CNN

Maybe we could increase our accuracy if we applied an FFT to the data before feeding it in?

2 of 20

Valence

Arousal

Dominance

3 of 20

Why emotion?

4 of 20

Applications

Clinical

Communication

Consumer Products

5 of 20

Methodology

6 of 20

Timeline

Video gathering

Data collection

(OpenBCI)

Model training

(CNN-LSTM)

7 of 20

Analysis

Real-time classification

Expansion

8 of 20

Video Gathering

Videos drawn from online databases

Catered to subject’s self-reported emotional profile

9 of 20

Data Collection

Cyton

OpenBCI

Headset

BT Dongle

OpenBCI

GUI

Python

Script

10 of 20

Data Labeling

Valence, Arousal, Dominance scores

Videos elicit emotional response

11 of 20

Classifiers

Valence

{0,1}

Arousal

{0,1}

Dominance

{0,1}

12 of 20

Classifier Architecture

Convolutional

Layer

ELU

Dropout

Batchnorm

Maxpool

LSTM

Layer

Dropout

Batchnorm

Fully connected

Softmax

13 of 20

Classifier Results

Epochs

Accuracy

Valence

Arousal

Dominance

Training

Validation

99.7%

89.9%

92.5%

Test acc:

14 of 20

Comparison in Binary VAD Classification

Group	Valence Accuracy	Arousal Accuracy	Dominance Accuracy
EE-JAMS (ours)	99.7%	89.9%	92.5%
M1M2 (Akter et al., 2022)	99.22%	97.80%	N/A
MMResLSTM (Ma et al., 2019)	92.87%	92.30%	N/A
Multi Column CNN (Jha et al., 2024)	86.23%	84.54%	85.02%

^*Other papers used DEAP/SEED datasets rather than just one participant

15 of 20

Discussion

16 of 20

Limitations

Further Directions

Emotions have no standardized biomarkers

Continuous classifier
Unsupervised learning
Identifying latent emotional variables

Corruptible signals (i.e. EMG)

Personalized AND generalized models

Accuracy is upper bounded by self-reported labels

Dataset augmentation & signal processing

17 of 20

Further Directions: BCI Cap

Discreet BCI wearable for day-to-day use

18 of 20

Demonstration

19 of 20

20 of 20

Thank you!

Questions they may ask:

What are other methods for emotion classification? Why just binary VAD classification?

Given our relatively small data set, we went with just binary classification, and this is indeed a limitation of our device

How does your model compare to other literature out there?

In 2016, papers achieve accuracy of 73.14% for valence and 73.06% for arousal using minimum-redundancy maximum-relevance and Support Vector Machine
Current state of the art methods achieve 99.89% accuracy in binary VAD classification. We are using only one subject whereas other papers typically use many subjects (32 for DEAP data set with 40 1 minute videos each, 15 for SEED data set). These papers will achieve classification very high accuracies on datasets such as DEAP of 99.22% using 2 seconds of EEG data but they of course have much more data.
We had in total about an hour and a half of video data

Why is it called EE-JAMS?

We originally