Topical Review
5
performance for each class [199]. If these conditions are not
met, the Kappa metric or the confusion matrix are more infor-
mative performance measures [66]. The sensitivity-specicity
pair, or precision, can be computed from the confusion matrix.
When the classication depends on a continuous parameter
(e.g. a threshold), the receiver operating characteristic (ROC)
curve, and the area under the curve (AUC) are often used.
Classier performance is generally computed ofine on
pre-recorded data, using a hold-out strategy: some datasets
are set aside to be used for the evaluation, and are not part of
the training dataset. However, some authors also report cross-
validation measures estimated on training data, which may
over-rate the performance.
The contribution of classier performance to overall BCI
performance strongly depends on the orchestration of the BCI
subcomponents. This orchestration is highly variable given the
variety of BCI systems (co-adaptive, hybrid, passive, self- or
system- paced). The reader is referred to [212] for a compre-
hensive review of evaluation strategies in such BCI contexts.
3. Past methods and current challenges
3.1. A brief overview of methods used ten years ago
In our original review of classication algorithms for EEG-
based BCIs published ten years ago, we identied ve main
families of classiers that had been explored: linear classi-
ers, neural networks, non-linear Bayesian classiers, nearest
neighbour classiers and classier combinations [141].
Linear classiers gather discriminant classiers that use
linear decision boundaries between the feature vectors of each
class. They include linear discriminant analysis (LDA), regu-
larized LDA and support vector machines (SVMs). Both LDA
and SVM were, and still are, the most popular types of clas-
siers for EEG based-BCIs, particularly for online and real-
time BCIs. The previous review highlighted that in terms of
performances, SVM often outperformed other classiers.
Neural networks (NN) are assemblies of articial neurons,
arranged in layers, which can be used to approximate any non-
linear decision boundary. The most common type of NN used
for BCI at that time was the multi-layer perceptron (MLP),
typically employing only one or two hidden layers. Other NN
types were explored more marginally, such as the Gaussian
classier NN or learning vector quantization (LVQ) NN.
Non-linear Bayesian classiers are classiers modeling the
probability distributions of each class and use Bayes’ rule to
select the class to assign to the current feature vector. Such
classiers notably include Bayes quadratic classiers and hid-
den Markov models (HMMs).
Nearest neighbour classiers assign a class to the cur-
rent feature vector according to its nearest neighbours. Such
neighbours could be training feature vectors or class proto-
types. Such classiers include the k-nearest neighbour (kNN)
algorithm or Mahalanobis distance classiers.
Finally, classier combinations are algorithms combin-
ing multiple classiers, either by combining their outputs
and/or by training them in ways that maximize their comple-
mentarity. Classier combinations used for BCI at the time
included boosting, voting or stacking combination algorithms.
Classier combination appeared to be amongst the best per-
forming classiers for EEG based BCIs, at least in ofine
evaluations.
3.2. Challenges faced by current EEG signal classication
methods
Ten years ago, most classiers explored for BCI were rather
standard classiers used in multiple machine learning
problems. Since then, research efforts have focused on iden-
tifying and designing classication methods dedicated to the
specicities of EEG-based BCIs. In particular, the main chal-
lenges faced by classication methods for BCI are the low
signal-to-noise ratio of EEG signals [172, 228], their non-
stationarity over time, within or between users, where same-
user EEG signals varying between or even within runs [56, 80,
109, 145, 164, 202], the limited amount of training data that is
generally available to calibrate the classiers [108, 137], and
the overall low reliability and performance of current BCIs
[109, 138, 139, 229].
Therefore, most of the algorithms studied these past ten
years aimed at addressing one or more of these challenges.
More precisely, adaptive classiers whose parameters are
incrementally updated online were developed to deal with
EEG non-stationarity in order to track changes in EEG prop-
erties over time. Adaptive classiers can also be used to deal
with limited training data by learning online, thus requir-
ing fewer ofine training data. Transfer learning techniques
aim at transferring features or classiers from one domain,
e.g. BCI subjects or sessions, to another domain, e.g. other
subjects or other sessions from the same subject. As such
they also aim at addressing within or between-subjects non-
stationarity and limited training data by complementing the
few training data available with data transferred from other
domains. Finally in order to compensate for the low EEG
signal-to-noise ratio and the poor reliability of current BCIs,
new methods were explored to process and classify signals in
a single step by merging feature extraction, feature selection
and classication. This was achieved by using matrix (notably
Riemannian methods) and tensor classiers as well as deep
learning. Additional methods explored were targeted speci-
cally at learning from limited amount of data and at dealing
with multiple class problems. We describe these new families
of methods in the following.
4. New EEG classication methods since 2007
4.1. Adaptive classiers
4.1.1. Principles.
Adaptive classiers are classiers whose
parameters, e.g. the weights attributed to each feature in a lin-
ear discriminant hyperplane, are incrementally re-estimated
and updated over time as new EEG data become available
[200, 202]. This enables the classier to track possibly chang-
ing feature distribution, and thus to remain effective even with
non-stationary signals such as an EEG. Adaptive classiers
for BCI were rst proposed in the mid-2000s, e.g. in [30, 72,
J. Neural Eng. 15 (2018) 031005
剩余28页未读，继续阅读
qmzqmz
- 粉丝: 0
- 资源: 2
我的内容管理
展开
最新资源
- 新型矿用本安直流稳压电源设计：双重保护电路
- 煤矿掘进工作面安全因素研究：结构方程模型
- 利用同位素位移探测原子内部新型力
- 钻锚机钻臂动力学仿真分析与优化
- 钻孔成像技术在巷道松动圈检测与支护设计中的应用
- 极化与非极化ep碰撞中J/ψ的Sivers与cos2φ效应：理论分析与COMPASS验证
- 新疆矿区1200m深孔钻探关键技术与实践
- 建筑行业事故预防：综合动态事故致因理论的应用
- 北斗卫星监测系统在电网塔形实时监控中的应用
- 煤层气羽状水平井数值模拟：交替隐式算法的应用
- 开放字符串T对偶与双空间坐标变换
- 煤矿瓦斯抽采半径测定新方法——瓦斯储量法
- 大倾角大采高工作面设备稳定与安全控制关键技术
- 超标违规背景下的热波动影响分析
- 中国煤矿选煤设计进展与挑战：历史、现状与未来发展
- 反演技术与RBF神经网络在移动机器人控制中的应用
资源上传下载、课程学习等过程中有任何疑问或建议，欢迎提出宝贵意见哦~我们会及时处理！
点击此处反馈
