convolutional neural networks deepa

Technical seminar on

Convolutional Neural Networks

forP300 Detection with Application to

Brain-Computer InterfacesDate: 21-03-2012

Presented By: Deepa D. Shedi 1KS08CS026 8th sem

CSE

OVERVIEW Introduction The P300 Speller Matrix and Detection Database Existing Systems Convolutional Neural Network Input Normalization Neural Network Topology Learning Classifiers Result for P300 Detection Network Analysis Character Recognition Rate Information Transfer Rate Discussion Conclusion

INTRODUCTIONBrain Computer Interface : A Brain-Computer Interface (BCI) is a specific type of human-computer interface that enables the direct communication between human and computers by analyzing brain measurements.

Eelectroencephalogram(EEG):It is a measure of brain’s voltage fluctuations as detected from scalp electrodes. It captures typical patterns of P300 signals.

P300 :EPRs are voltage fluctuations in the EEG induced with in the brain that are time locked to sensory or motor events.The P300 is positive bump in the ERP named so because it starts about 300 milliseconds after an event. 1

THE P300 SPELLER MATRIX AND DETECTION

The two classification problems. P300 detection. Character Recognition.

Fig 1: P300 detection Fig 2 : Character recognition.

2

DATABASEData set contains a complete record of P300 evoked potentials from two subjects.

The subject was presented with a matrix of size 6X6.

2 out of 12 intensifications for rows/columns.

The number of samples for both databases and for each subject is presented.

Table 1: Database Size for Each Subject3

EXISTING SYSTEMS

This section describes some of the best techniques that have been proposed during the III BCI competition.

Support Vector Machine(SVM)

Band-pass Filtering

Frequency Filtering And Principal Component Analysis (PCA)

Gradient Boosting Method

Component Classifier Linear Discriminant Analysis (LDA)

4

CONVOLUTIONAL NEURAL NETWORK

The classifiers that are used for the detection of P300 responses are based on a convolutional neural network(CNN).

Neural network is a multilayer perceptron (MLP).

Neural network is used for object recognition and handwriting character recognition.

A classifier based on a CNN seems to be a good approach for EEG classification

5

INPUT NORMALIZATION

Fig. 3. Electrode map.

Steps in Normalization:

Step 1: Subsampling of EEG signal to reduce the size of the data to analyze and divided by two.

Step 2: Bandpass filtering of signal to keeponly relevant frequencies.

Normalized Signal :

6

NEURAL NETWORK TOPOLOGY

Fig. 4. Neural Network Architecture.

7

CONTD.The network topology is described as follows:

L0: The input layer. Ii,j with 0 ≤ i < Nelec and 0 ≤ j < Nt.

L1: The first hidden layer is composed of Ns maps. We define L1Mm, the map number m. Each map of L1 has the size Nt.

L2: The second hidden layer is composed of 5Ns maps. Each map of L2 has six neurons.

L3: The third hidden layer is composed of one map of 100 neurons. This map is fully connected to the different maps of L2.

L4: The output layer. This layer has only one map of two neurons, which represents the two classes of the problem (P300 and no P300). This layer is fully connected to L3. 8

LEARNINGA neuron in the network is defined by n(l, m, j).This sigmoid function for only one map in the layer:

Convolution of the input signal for first two hidden layers:

The classical sigmoid function is used for the twolast layers:

9

LEARNING σm

1(j) represents the scalar product between a set of inputneurons and the weight connections between these neurons and the neuron number j in the map m in the layer l.

σm1(j) for the four layers.

For L2 :

For L1 :

This layer aims at finding the best electrodes combination for the classification.

This layer translates subsampling and temporal filters.

10

LEARNINGFor L3 :

In this layer, each neuron has NsNf +1 input weights. L3 contains 100(5*6*Ns) input connections.

For L4 :

Each neuron of L4 is connected to each neuron of L3.

11

LEARNINGLearning rate ϒ for layers L1 and L2 :

Learning rate ϒ for layers L1 and L2 :

The detection of a P300 wave :

12

CLASSIFIERSSingle classifiers:

• CNN-1

• CNN-2a

• CNN-2b

• CNN-3

Multiclassifiers :

• MCNN-1

• MCNN-2

• MCNN-313

RESULT FOR P300 DETECTIONResults of the P300 Detection for Subject A :

Results of the P300 Detection for Subject B :

Table 3: Results Obtained After the P300 detection of Subject B.

Table 2: Results Obtained After the P300 detection of Subject A.

14

RESULT FOR P300 DETECTIONMeasures for evaluating the quality of results :

Subject B allows getting better results for the classification.

15

NETWORK ANALYSISSpatial filters obtained with CNN-1 for subject A.

Spatial filters obtained with CNN-1 for subject B.

16

Fig. 5. Spatial filters obtained with CNN-1 for subject A.

Fig.6. Spatial filters obtained with CNN-1 for subject B.

NETWORK ANALYSISElectrode Ranking :

The coordinate of the character are defined by :

Cumulated probabilities of the P300 detection:Table 4: Ranking of Electrodes

17

CHARACTER RECOGNITION RATE

Table 5:Character Recognition Rate (in Percent) for the Different Classifiers 18

INFORMATION TRANSFER RATE

Information Transfer Rate :

T is the time needed to recognize one character. T is defined by :

Best recognition rate when only 10 epochs are used.

19

INFORMATION TRANSFER RATE

Fig. 7. ITR (in bits per minute) in relation to the number of epochs.

20

DISCUSSIONThe database has two main interests.

• It forces the system to reach the limit of the P300 detection.

• It is an excellent challenge for the machine learning community.

Steady-state visual evoked potentials (SSVEPs), the user has to focus on some visual Stimuli.

The interest of convolutional neural networks is double.• It allows a high performance in the classification.• They can allow deeper analysis of brain activity.

21

DISCUSSIONIn the P300 speller, it is possible that the subject may not have always focused on the expected target.

Table 6:Confusion of Character Recognition

Table 7:Comparison of the Recognition Rate and the ITR with Other Results.

Comparison

22

CONCLUSIONThis model is based on a convolutional neural network.

It outperforms the best method in two situations: • When the number of electrodes is restricted to 8.• when the number of considered epochs is 10.

The recall of the P300 detection is the main feature that dictates the overall performance of the P300 speller.

The detection of P300 waves remains a very challenging problem for both the machine learning and neuroscience communities.

23

REFERENCESC.W. Anderson, S.V. Devulapalli, and E.A. Stolz, “Determining Mental State from EEG Signals Using Parallel Implementations of Neural Networks,” Proc. IEEE Workshop Neural Networks for Signal in Processing, pp. 475-483, 1995.

K. Chellapilla, S. Puri, and P.Y. Simard, “High Performance Convolutional Neural Networks for Document Processing,” Proc.10th Int’l Workshop Frontiers in Handwriting Recognition, 2006.

B. Blankertz, K.-R. Muller, G. Curio, T.M. Vaughan, G. Schalk, J.R.Wolpaw, A. Schlogl, C. Neuper, G. Pfurtscheller, T. Hinterberger, M. Schroder, and N. Birbaumer, “The BCI Competition 2003: Progress and Perspectives in Detection and Discrimination of EEG Single Trials,” IEEE Trans. Biomedical Eng., vol. 51, no. 6, pp. 1044-1051, June 2004.

D.J. Krusienski, E.W. Sellers, D. McFarland,T.M. Vaughan, and J.R. Wolpaw, “Toward Enhanced P300 Speller Performance,” J. Neuroscience Methods, vol. 167, pp. 15-21, 2008.

G. Schalk, D.J. McFarland, T. Hinterberger, N. Birbaumer, and J. Wolpaw, “BCI2000: A General-Purpose Brain-Computer Interface(BCI) System,” IEEE Trans. Biomedical Eng., vol. 51, no. 6, pp. 1034-1043, June 2004.

27

Thank You!!

Questions???