CNN for P300 Detection

Convolutional Neural Networks for P300 Detection with Application to  Brain-Computer Interfaces》

The paper is organized as follows: The P300 wave, the oddball paradigm, and the database are presented in Section 2. The neural network is described in Section 3. Section 4  describes the different classifiers. Finally, the results and their discussion are detailed in Sections 5 and 6.

The P300 wave is an event-related potential (ERP) which can be recorded via EEG.

1.Seven classifiers based on the CNN are proposed:  four single classifiers with different features set and three multiclassifiers.

event-related potentials(ERPs),事件相关诱发电位，conventional means，传统的方法，motor disabilities，运动障碍，noninvasive，非侵害的，无创

2.The proposed approach provides also a new way for analyzing brain activities due to  the receptive field of the CNN models.

基于感受野提出了神经认知机（neocognitron），可看作是CNN的第一个实现网络。

3.A BCI is usually decomposed into  four main parts that translate the neural signal processing. 

4.The EEG classification strategy depends on the stimulus and, thereby, the response to detect: event-related potentials, steady-state evoked potentials, motor imagery, or slow cortical potentials. The

expected EEG drives the classification to some specific  feature extraction methods.

5.Pattern recognition techniques are used for the classification and the detection of specific brain signals.Most of the effective solutions use machine learning models[5-8].Although neuroscience provides knowledge and guidelines about how to process and detect the expected signals,machine learning techniques allow modeling the signal variability over time and over subjects. Neural networks[9],[10],[11],[12],[13],[14],support vector machines( SVMs)[15],[16],and  hidden Markov models[17],[18]have already been applied to BCI and  EEG classification.Neural networks using backpropagation were used for the first time for readiness potential pattern recognitionin[19],proving that neural networks can be used for classifying EEG and for tailoring a brain machine interface.

6.Also,one interesting property of CNN models is  the semantic of the weights once the network is trained. The receptive field/convolution kernel can be easily interpreted and can provide a diagnostic about the type of high-level features to detect.We propose using CNN models and their combination, for the first time, for the detection of P300 waves.

the semantic of the weights ？？

7.The P300 wave is an event-related potential (ERP) which can be recorded via EEG. TIn a P300 speller, the main goal is to detect  the P300 peaks in the EEG accurately and instantly. 

8.The detection of P300 responses. It corresponds to a  binary classification: One class represents signals

that correspond to a P300 wave,the second class is oppsite.For this classification problem, the

creation of  the ground truth can be quite challenging.

the ground truth??  

9.In these experiments, the subject was  presented with  a matrix of size 6*6. Each cell of the matrix

contains a character: [A-Z], [1-9], and [_].      Thus,  2 out of 12 intensifications of rows/columns

highlighted the expected character,

2 out of 12 ?? 

10.Therefore, 30 possible P300 responses should be detected for the character recognition.

11.This section describes some of the best techniques that have been proposed during the third BCI competition. They also correspond to the state of the art for the P300 speller.

<1>.The solution proposed by Rakotomamonjy and Guigue [16] is based on an ensemble of  SVMs.

<2>.First, data are preprocessing with bandpass filtering at 0.5-8 Hz. Then, eye movement artifacts are removed by using independent component analysis( ICA) for the whole data set.The classification is based on  SVMs[29]

<3>. xxx is based on frequency filtering and principal component analysis ( PCA) for the preprocessing steps. The feature selection uses  t-statistic values at each data point in each channel.

<4>.uses a  gradient boosting method

<5>.uses bagging with component classifier  linear discriminant analysis ( LDA) [29].For the input, the signal is bandpass filtered between 0.5 and 15 Hz and 10 channels are selected before the classification.

12.CNN.This type of neural network is a multilayer perceptron ( MLP) with a special topology and contains more than one hidden layer.It allows  automatic features extraction within its layers

and it keeps as input the raw information without specific normalization, except for scaling and centering the input vector. 

One interest on convolutional neural network is the possibility of including, inside the network, high-level knowledge that is directly related to the problem, contrary to kernel-basedmethods [20].

However, the topology of the network remains  an empirical choice and depends on the application. The topology translates different successive signal/neural processing steps.

an empirical choice，根据经验的选择，successive 连续的，

The interest of the CNN is to directly classify the raw signal and to integrate the signal processing functions within the  discriminant steps.

discriminant ，判别式，判别分析，

It is better to let the network  extract  the most discriminant features by constructing high-level features throughout the propagation step.

疑问：为什么都是0.1Hz到20Hz的滤波？？

13.The input of the CNN is a matrix Nelec * Nt, where Nt is the number of points that are considered for the analysis: Nt - SF*TS. Nt corresponds to the number of recorded samples in T Ss with the sampling rate SF.

14.The network topology is the key feature in the classifier. The network is composed of  five layers, which are themselves composed of one or several maps. A map represents a layer entity, which has a specific semantic: Each map of  the first hidden layer is a channel combination. The second hidden

layer subsamples and transforms the signal in the time domain. The classifier architecture is presented in Fig. 3.（看pdf版） The number of neurons for each map is presented between brackets; the size of the convolution kernel is between hooks.

brackets，方括号

In the proposed strategy,  the kernels are vectors and not matrix, like in CNNs for image recognition. The reason is to not mix in one kernel features related to the space and time domain.

15.The network topology is described as follows:   具体看pdf版

. L0: The input layer. Ii;j with 0 i

.  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.

16.Learning，具体的学习过程，一大堆的数据公式，，每一个layer写的非常非常详细，

17.CLASSIFIERs，

We present here seven classifiers based on the convolutional neural network that was persented in the previous section.This classifiers will be used as the core of the different models.Among the presented classifiers, CNN-1, CNN-2, and CNN-3 are single classifiers whereas MCNN-1, MCNN-2, MCNN-3 are based on a multiclassifiers strategy,like most of the efficient methods that achieve good results on P300 detection.

is identical to, 等同于，但不是完全一样，

MCNN-1 is composed of  five classifiers. Each classifier is trained on a different database. Each training database contains all the P300 patterns and a fifth of the non-P300 patterns from the main training database.

疑问，为什么通篇还串讲着“字符识别”？？

首先搞清楚文章中所说的database的数据格式是什么。

Therefore, the data from the third BCI competition are noteworthy primarily because they present an excellent challenge.     BCI比赛的数据的格式是什么？？

Bagging Predictors，其中bagging指什么？？

关于文章中的疑问的回答：

1. 机器视觉的论文中处理图像时经常提到ground truth代表什么意思呢？

答：标定好的真实数据！

2. 机器学习里经常出现ground truth这个词，能否准确解释一下？

就是参考标准，一般用来做error quantification。比方说要根据历史数据预测某一时间的温度，ground truth就是那个时间的真实温度。error就是(predicted temperature - real temprature)。

Ground truth当然还可以用来做reinforcement learning，就是在学习中加入奖励机制。比方说程序的输出越接近ground truth，用来产生这个结果的数据的weight越大。

Wiki中的解释是：

In machine learning, the term "ground truth" refers to the accuracy of the training set's classification for supervised learning techniques. This is used in statistical models to prove or disprove researchhypotheses. The verb "ground truthing" refers to the process of gathering the proper objective data for this test. Compare with gold standard (test).

Bayesian spam filtering is a common example of supervised learning. In this system, the algorithm is manually taught the differences between spam and non-spam. This depends on the ground truth of the messages used to train the algorithm; inaccuracies in that ground truth will correlate to inaccuracies in the resulting spam/non-spam verdicts.

3.为什么说bagging是减少variance，而boosting是减少bias?

这个应该最直接的理解：

boosting是把许多弱的分类器组合成一个强的分类器。弱的分类器bias高，而强的分类器bias低，所以说boosting起到了降低bias的作用。variance不是boosting的主要考虑因素。

bagging是对许多强（甚至过强）的分类器求平均。在这里，每个单独的分类器的bias都是低的，平均之后bias依然低；而每个单独的分类器都强到可能产生overfitting的程度，也就是variance高，求平均的操作起到的作用就是降低这个variance。

具体参考：http://www.zhihu.com/question/26760839

一篇中文博客：深度学习与脑机接口_1（基于卷积神经网络的P300信号检测）

http://www.cnblogs.com/lijingcong/p/3732982.html