Limits...
A Generalizable Brain-Computer Interface (BCI) Using Machine Learning for Feature Discovery.

Nurse ES, Karoly PJ, Grayden DB, Freestone DR - PLoS ONE (2015)

Bottom Line: The classifier does not use extensive a-priori information, resulting in reduced reliance on highly specific domain knowledge.Instead of pre-defining features, the time-domain signal is input to a population of multi-layer perceptrons (MLPs) in order to perform a stochastic search for the best structure.Our new approach has been shown to give accurate results across different motor tasks and signal types as well as between subjects.

View Article: PubMed Central - PubMed

Affiliation: NeuroEngineering Laboratory, Department of Electrical & Electronic Engineering, The University of Melbourne, Melbourne, VIC, Australia, 3010; Centre for Neural Engineering, The University of Melbourne, Melbourne, VIC, Australia, 3010.

ABSTRACT
This work describes a generalized method for classifying motor-related neural signals for a brain-computer interface (BCI), based on a stochastic machine learning method. The method differs from the various feature extraction and selection techniques employed in many other BCI systems. The classifier does not use extensive a-priori information, resulting in reduced reliance on highly specific domain knowledge. Instead of pre-defining features, the time-domain signal is input to a population of multi-layer perceptrons (MLPs) in order to perform a stochastic search for the best structure. The results showed that the average performance of the new algorithm outperformed other published methods using the Berlin BCI IV (2008) competition dataset and was comparable to the best results in the Berlin BCI II (2002-3) competition dataset. The new method was also applied to electroencephalography (EEG) data recorded from five subjects undertaking a hand squeeze task and demonstrated high levels of accuracy with a mean classification accuracy of 78.9% after five-fold cross-validation. Our new approach has been shown to give accurate results across different motor tasks and signal types as well as between subjects.

No MeSH data available.


Related in: MedlinePlus

Spatial filtering of artificial neural networks (ANNs).The absolute weight values between the input and first hidden layer are averaged across time and over neurons in the first hidden layer to obtain a single value for each EEG electrode channel. The weights are then normalized so the heat map shows the minimum weight at 0 and the maximum weight at 1. The figure for each participant was obtained from the average weights over the five-folds of cross-validation A Subject A used 62 electrode channels. B Subject B used 62 electrode channels. C Subject C used 27 electrode channels. D Subject D used 46 electrode channels. E Subject E used 10 electrode channels.
© Copyright Policy
Related In: Results  -  Collection

License
getmorefigures.php?uid=PMC4482677&req=5

pone.0131328.g009: Spatial filtering of artificial neural networks (ANNs).The absolute weight values between the input and first hidden layer are averaged across time and over neurons in the first hidden layer to obtain a single value for each EEG electrode channel. The weights are then normalized so the heat map shows the minimum weight at 0 and the maximum weight at 1. The figure for each participant was obtained from the average weights over the five-folds of cross-validation A Subject A used 62 electrode channels. B Subject B used 62 electrode channels. C Subject C used 27 electrode channels. D Subject D used 46 electrode channels. E Subject E used 10 electrode channels.

Mentions: Through analysis of network input layer weights, the influence of various spatial and temporal patterns can be inferred. Fig 9 shows the network input weights over all the cross validation partitions. The heat maps in Fig 9 are a mapping from the input layer to the first hidden layer of the network (see Fig 1) and thus represent a small part of the true interactions between the input signal and the output. The weights for each electrode channel were obtained by averaging the absolute input weights over time and across the number of neurons in the next layer. In this way, an implicit measure of the relative importance of each channel is obtained. For Fig 9A, 9B and 9D, the largest average weights were from the central and frontal-central group electrodes, which were placed above the motor cortex [23, 24]. In other words, the neural network classifiers extracted features with the greatest weighting from the signal sources above the motor cortex and pre-motor cortex [23, 24]. The same pattern is less evident in Fig 9C and 9E, as these participants used fewer electrode channels so the difference in spatial weighting is less pronounced.


A Generalizable Brain-Computer Interface (BCI) Using Machine Learning for Feature Discovery.

Nurse ES, Karoly PJ, Grayden DB, Freestone DR - PLoS ONE (2015)

Spatial filtering of artificial neural networks (ANNs).The absolute weight values between the input and first hidden layer are averaged across time and over neurons in the first hidden layer to obtain a single value for each EEG electrode channel. The weights are then normalized so the heat map shows the minimum weight at 0 and the maximum weight at 1. The figure for each participant was obtained from the average weights over the five-folds of cross-validation A Subject A used 62 electrode channels. B Subject B used 62 electrode channels. C Subject C used 27 electrode channels. D Subject D used 46 electrode channels. E Subject E used 10 electrode channels.
© Copyright Policy
Related In: Results  -  Collection

License
Show All Figures
getmorefigures.php?uid=PMC4482677&req=5

pone.0131328.g009: Spatial filtering of artificial neural networks (ANNs).The absolute weight values between the input and first hidden layer are averaged across time and over neurons in the first hidden layer to obtain a single value for each EEG electrode channel. The weights are then normalized so the heat map shows the minimum weight at 0 and the maximum weight at 1. The figure for each participant was obtained from the average weights over the five-folds of cross-validation A Subject A used 62 electrode channels. B Subject B used 62 electrode channels. C Subject C used 27 electrode channels. D Subject D used 46 electrode channels. E Subject E used 10 electrode channels.
Mentions: Through analysis of network input layer weights, the influence of various spatial and temporal patterns can be inferred. Fig 9 shows the network input weights over all the cross validation partitions. The heat maps in Fig 9 are a mapping from the input layer to the first hidden layer of the network (see Fig 1) and thus represent a small part of the true interactions between the input signal and the output. The weights for each electrode channel were obtained by averaging the absolute input weights over time and across the number of neurons in the next layer. In this way, an implicit measure of the relative importance of each channel is obtained. For Fig 9A, 9B and 9D, the largest average weights were from the central and frontal-central group electrodes, which were placed above the motor cortex [23, 24]. In other words, the neural network classifiers extracted features with the greatest weighting from the signal sources above the motor cortex and pre-motor cortex [23, 24]. The same pattern is less evident in Fig 9C and 9E, as these participants used fewer electrode channels so the difference in spatial weighting is less pronounced.

Bottom Line: The classifier does not use extensive a-priori information, resulting in reduced reliance on highly specific domain knowledge.Instead of pre-defining features, the time-domain signal is input to a population of multi-layer perceptrons (MLPs) in order to perform a stochastic search for the best structure.Our new approach has been shown to give accurate results across different motor tasks and signal types as well as between subjects.

View Article: PubMed Central - PubMed

Affiliation: NeuroEngineering Laboratory, Department of Electrical & Electronic Engineering, The University of Melbourne, Melbourne, VIC, Australia, 3010; Centre for Neural Engineering, The University of Melbourne, Melbourne, VIC, Australia, 3010.

ABSTRACT
This work describes a generalized method for classifying motor-related neural signals for a brain-computer interface (BCI), based on a stochastic machine learning method. The method differs from the various feature extraction and selection techniques employed in many other BCI systems. The classifier does not use extensive a-priori information, resulting in reduced reliance on highly specific domain knowledge. Instead of pre-defining features, the time-domain signal is input to a population of multi-layer perceptrons (MLPs) in order to perform a stochastic search for the best structure. The results showed that the average performance of the new algorithm outperformed other published methods using the Berlin BCI IV (2008) competition dataset and was comparable to the best results in the Berlin BCI II (2002-3) competition dataset. The new method was also applied to electroencephalography (EEG) data recorded from five subjects undertaking a hand squeeze task and demonstrated high levels of accuracy with a mean classification accuracy of 78.9% after five-fold cross-validation. Our new approach has been shown to give accurate results across different motor tasks and signal types as well as between subjects.

No MeSH data available.


Related in: MedlinePlus