This document describes a study that used convolutional neural networks (CNNs) to improve classification accuracy in a tactile P300-based brain-computer interface (BCI). The researchers designed CNN input volumes from EEG signal intervals and implemented a LeNet-inspired CNN architecture. Classification accuracy improved from around 80% with raw EEG data to 100% when using moving averages. The CNN approach allowed for non-personalized training, demonstrating the potential of CNNs to advance tactile P300 BCI accuracy and usability.
Similar to Convolutional Neural Network Architecture and Input Volume Matrix Design for ERP Classifications in a Tactile P300–based Brain–Computer Interface
Similar to Convolutional Neural Network Architecture and Input Volume Matrix Design for ERP Classifications in a Tactile P300–based Brain–Computer Interface (20)
Convolutional Neural Network Architecture and Input Volume Matrix Design for ERP Classifications in a Tactile P300–based Brain–Computer Interface
1. Convolutional Neural Network
Architecture and Input Volume Matrix
Design for ERP Classifications in a Tactile
P300-Based Brain-Computer Interface
1
Convolutional Neural Network
Architecture and Input Volume Matrix
Design for ERP Classifications in a Tactile
P300-Based Brain-Computer Interface
1
Takumi Kodama and Shoji Makino
Life Science Center of TARA, University of Tsukuba,
Tsukuba, Japan
2. Introduction - What’s the BCI?
● Brain Computer Interface (BCI)
○ Exploits user intentions ONLY using brain responses
2
3. Introduction - P300-based BCI
1, Stimulate touch sensories 2, Classify brain response
A
B
Target
Non-Target
P300 brainwave response
3
● Tactile (Touch) P300-based BCI paradigm
○ Predict user’s intentions by decoding P300 responses
○ Strong peak could be aroused by vibrotactile stimuli
BCI user
A
B
4. Introduction - Our Approach
4
● Full-body Tactile P300-based BCI (fbBCI) [1]
○ Applies six vibrotactile stimulus patterns to user’s back
○ User can take experiment with their body lying down
[1] T. Kodama, S. Makino and T.M. Rutkowski, “Tactile Brain-Computer Interface Using Classification of P300 Responses Evoked by Full Body
Spatial Vibrotactile Stimuli,” in Proc. the Asia-Pacific Signal and Information Processing Association Annual Summit and Conference 2016 (APSIPA
ASC 2016), IEEE Press, pp. Article ID: 176, Dec. 2016.
5. [2] T. Kodama, K. Shimizu, S. Makino and T.M. Rutkowski, “Full–body Tactile P300–based Brain–computer Interface Accuracy Refinement,” in
Proc. the International Conference on Bio-engineering for Smart Technologies 2016 (BioSMART 2016), IEEE Press, pp. 20–23, Dec. 2016.
Introduction - fbBCI exp. results
5
● Classification accuracies with personal trainings [2]
○ SWLDA
■ Average result: 57.48 %
○ Linear SVM:
■ Average result: 58.5 %
○ Non-Linear SVM:
■ Average result: 59.83 %
6. [2] T. Kodama, K. Shimizu, S. Makino and T.M. Rutkowski, “Full–body Tactile P300–based Brain–computer Interface Accuracy Refinement,” in
Proc. the International Conference on Bio-engineering for Smart Technologies 2016 (BioSMART 2016), IEEE Press, pp. 20–23, Dec. 2016.
Introduction - fbBCI results (2)
6
● Classification accuracies with personal trainings [2]
○ SWLDA
■ Average result: 57.48 %
○ Linear SVM:
■ Average result: 58.5 %
○ Non-Linear SVM:
■ Average result: 59.83 %
1. Classification accuracies were not enough
for a practical usage of BCI
2. Requires user-specific classifier models for
each user
Problems
7. 1. Confirm an effectiveness of the fbBCI modality by
improving stimulus pattern classification accuracies
2. Achievement of non-personal-training ERP
classifications using a neural network model
Introduction - Research Purpose
7
8. ● Convolutional Neural Networks (CNN) [3]
○ Pixel elements were convolved with filters in layers
○ Neural networks were applied to the output vectors
Method
[3] Y. LeCun, L. Bottou, Y. Bengio, and P. Haffner, “Gradient-based learning applied to document recognition,” Proceedings of the IEEE, vol. 86, no.
11, pp. 2278–2324, 1998.
Input volume design CNN architecture
9. ● Convolutional Neural Networks (CNN) [3]
○ Pixel elements were convolved with filters in layers
○ Neural networks were applied to the output vectors
Method
Input volume design CNN architecture
[3] Y. LeCun, L. Bottou, Y. Bengio, and P. Haffner, “Gradient-based learning applied to document recognition,” Proceedings of the IEEE, vol. 86, no.
11, pp. 2278–2324, 1998.
10. ● Transform ERP intervals to feature vectors
Ch○○
Method - Input volume design
1. Captures 800 ms long after
vibrotactile stimulus onsets
2. Converts to feature vectors
with their potentials
11. Method - Input volume design
11
3. Feature vectors were
deployed in a 20 × 20
squared matrix
4. Matrices generated in
each electrode channel
and mean of all electrodes
were concatenated into a 3
× 3 grid input volume
● Transform feature vectors to input volumes
12. ● Convolutional Neural Networks (CNN) [3]
○ Pixel elements were convolved with filters in layers
○ Neural networks were applied to the output vectors
Method
[3] Y. LeCun, L. Bottou, Y. Bengio, and P. Haffner, “Gradient-based learning applied to document recognition,” Proceedings of the IEEE, vol. 86, no.
11, pp. 2278–2324, 1998.
Input volume design CNN architecture
13. Method - CNN architecture
● Overview of CNN architecture in fbBCI
○ CONV > POOL > CONV > POOL (LeNet)
○ (Ix, Iy) … Size of the input volume
○ (Ax, Ay) … Size of activation maps
13
MLP
15. Method - Non-personal-trainings
15
User 1
1
2 3 4
7 8 9
Classifier model
trained by user 2~10
ERP classification
● Evaluate with the classifier model which trained by other
nine participated user
5 6
10
16. Method - Non-personal-trainings
16
User 1
1
2 3 4
7 8 9
Classifier model
trained by user 2~10
ERP classification
● Evaluate with the classifier model which trained by other
nine participated user
5 6
10
User 10
10
1 2 3
6 7 8
Classifier model
trained by user 1~9
ERP classification
4 5
9
19. ● The fbBCI classification accuracy was dramatically
improved with CNN classifier model
○ 79.66 % with non averaged ERP intervals
○ 100 % with moving averaged ERP intervals
● The potential validity of fbBCI modality was reconfirmed
● A non–personal–training ERP classification was achieved
by CNN classifier model with high performance results
● In the future study, to implement the proposed methods for
the online environment would be the primary task
Conclusions
19
24. ALS Patients
● Amyotrophic lateral sclerosis (ALS) patients
○ Have difficulty to move their muscle by themselves
○ BCI could be a communicating tool for them
24
…
…!
25. ● Grand mean ERP intervals in each electrode channel [1]
fbBCI ERP interval results
25
*Gray-shaded area … significant difference (p < 0.01) between targets and non-targets
26. 26
● ERP epoch averaging
○ To cancel background noise
Non averaged ERP Moving averaged ERP
ERP averaging