1. Convolutional Neural Network for Multi-Category Rapid Serial Visual Presentation BCI
[Ran Manor and Amir B.Geva]
Yehu Sapir
Outlines
Review
EEG/ERP (Event-Related Potential)
P300
Spatio-temporal data
CNN Network + regularization
Results and discussion

2. Review Categories of electrical cortical activities in Brain Computer Interface (BCI)μ β
μ (8–12 Hz) and β (12–30 Hz) rhythms.
Response of motor output of the brain.
P300 Evoked Potential (EP). dementia / Alzheimer
dementia / Alzheimer
Auditory, visual, or somatosensory Event-Related Potentials (ERP) (1–30 μV).
300–500 ms post-stimulus
Visual N100 and P200, ERPs with short latency
Response of the brain to a rapid visual stimulus.
Steady-State Visual Evoked Potentials (SSVEP).
Visual stimulations ( Hz) generates brain responses at the same freq.
Slow Cortical Potentials (SCP).
slow potential variations 0.5–10 sec after presenting stimulus
users control these potentials and use them to control the movement of a cursor on screen
movement
reduced cortical activation.

3. Review P300 signal P300 ERP at electrode Pz,
P300 = target detection for
RSVP (Rapid Serial Visual Presentation) tasks.
The amplitude and latency have a large variance
Average
of multiple
single trials
Amplitude Pz
Single trial response
Time [ms]

4. Subjects & Stimuli Aim Subjects Stimuli
classify ERPs into target and non-target images.
Subjects
15 subjects participated in a RSVP experiment,
8 females, 7 males with
age 25±5 years.
Stimuli
360 × 360 pixels, gray-scale images.
The images have the same mean luminance and contrast.
5 different categories including 145 exemplars each of faces, cars, painted eggs, watches, and planes.
one category
is the target

5. Experimental Procedure Brain computer interface (BCI)
Images presented in 4 blocks,
6525 images in each block, presented every 90–110ms (~10Hz).
20% of the images were targets, randomly distributed within each block.
different target category in each block
watches were not used as targets.
Use same blocks order across subjects.
Subjects should count targets.
Pause every 80–120 image
Reporte the Number of target.
Restart the count.

6. EEG Acquisition and Preprocessing
64 electrodes + 7 additional electrodes (to reject trials when blinking).
256Hz sampling rate
51Hz online LPF to prevent aliasing of high frequencies and remove powerline noise.
0.1Hz offline high-pass filter to remove slow drifts (DC).
Data segmentation: 1sec event-related segments (-100ms to +900ms of image presentation)
Single trial =1 Image
256 Samples
64 channels
(electrodes)
Down-sampling,
Normalizing &
Removing DC 64
target 64
µ=0,σ=1
+900ms
-100ms
µ=0,σ=1

7. Neural Network Architecture
Implementation
on Caffe and GPU
Time Samples
ReLU
ReLU
Dropout
ReLU
Dropout
ReLU
ReLU
softmax
channels
# output filters

8. Spatio-Temporal Regularizer
penalty term :
Where: conv. output l at time t
- regularization coefficient ,
Where: a matrix of zeros with 1 in position i, j
Regularized No-regularization

9. Parameters updates Train the Net. By Minimize the loss function
where Nsamples = training samples,
x(i) = i training sample,
y(i) = true label of sample i
hk = NN output unit k
Minimized with SGD,
Momentum update: V= µ*V + η*dW
Weights updates: W=W-V
Learning rate (η) = 0.001
Momentum coefficient (µ) = 0.9
network parameters were chosen empirically by cross-validation.

10. Performance Correct is distorted, due to imbalanced classes
positives +
negatives
random cross-validation procedure
80% training
20% testing (4 blocks per subject)
average of 10 runs
(true positive +
true negative) /2
true positive
false positive
Area Under the Curve
Correct is distorted,
due to imbalanced classes
Mean across subjects

11. Compare to Spatially Weighted FLD-PCA (SWFP)

12. Features Analysis
Representative spatial filters and their temporal features from a sample subject (504).
Spatial maps are similar to the spatial distribution of the P300 with a high amplitude at central-parietal electrodes.
spatial filters
distributed on the scalp
mean temporal activations
of the spatial filter

13. Sample of spatial filters and temporal activations for subjects
501
502
503
507

14. Sample of spatial filters and temporal activations for subjects
508
511
512
513

15. BCI Competition 2004 Benchmark
Compare our CNN model to BCI Competition, 2004
P300 Speller task.
Two subjects (A, B) F E D C B A L K J I H G R Q P O N M V U T S
false alarm rate %
hit rate
correct classification %
Subject
27.36
68.13
71.9
A [BCI Competition, 2004] 29 67
70.5
A [Cecotti and Gräser, 2011]
19.11 70
79.1
B [BCI Competition, 2004] 20 69
78.19
B [Cecotti and Gräser, 2011]

16. Cross-session performance
correctly
classified
(true positive +
true negative) /2
true positive
false positive

17. Thank You!