The task of a deep neural network (DNN) as a component of a brain-computer interface (BCI) is to analyze the measured EEG data and to recognize the neural signal patterns characteristic of a given motor movement. Our studies are intended to investigate the use of such a DNN in continuous control applications where the EEG signals need to be interpreted continuously, as well as to gain insights into the learned neural patterns. Method: We examined EEGNet, a commonly referenced Convolutional Neural Net (CNN) trained with trials from the BCI Competition IV 2a EEG data set. In addition to the impact of the size and temporal location of the trials on classification accuracy, we examined the contributions and temporal behavior of known neural patterns and their impact on system's response time and the amount of time for which the mental state was detected. Results: Because the optimal temporal positions of the trials are different for the neural patterns involved, we introduced cropped training in which the DNN is trained using trials with different temporal positions. This enabled the DNN to learn the neural patterns in the 0-8 Hz frequency range that are important for a short response time and the patterns in the 8-30 Hz frequency range that are important for determining state duration. Conclusion: Cropped training is essential for achieving a good response time, which could be improved by ~200 ms, as well as for a good state duration, which could be increased from 150 ms to 1.75 s.

Throughout the range of the Scarlet Rosefinch, its territorial song consists of 3–9 (usually 4–5) elements, of which there are 5 different types. The differences lie in the way the pitch of the element changes in time (frequency “slope”) and the width of the frequency band. Within a given type of song, the various elements can be present in almost any combination. Therefore, so many song types can be formed that the songs in even small parts of the species’ area are clearly distinct from one another. Despite this capacity for variation, however, by chance identical songs may be sung in widely separated parts of the area, in some cases by different subspecies.
 The species has not developed large‐scale dialects or regiolects based on a song tradition acquired during an early imprinting phase. Scarlet Rosefinches tend to breed in small colonies, groups of up to about 15 pairs characterized by the same type of song (song neighbourhoods, formed by the development of a microlect).
 Microlects develop by a founder effect. When males, near one‐year old or older, join one another to form isolated colonies after arrival in the breeding region, they adopt (“learn”) the song type that will eventually characterize the colony from the first male to arrive at the site. After the colony has been founded, in most cases each male uses only one type of song during a breeding season, with practically no variation of the temporal and frequency parameters.
 Singing the same type of song, the members of a colony accept one another sufficiently to allow the breeding territories to be closely packed. It appears that a long‐lasting capacity for acoustic learning, in combination with colony‐like breeding and great ecological flexibility, has allowed the Scarlet Rosefinch to become the most successful species of the genus Carpodacus.

<p>In this paper we report on our investigations on the use of a Deep Learning based brain-computer interface in the context of a continuous control application. A continuous control application is an application where a Deep neural network (DNN) is supposed to recognize defined motor-imagery related neural states from continuously measured EEG data stream and to initiate corresponding actions. Decisive for the quality of such an application is the achieved response time of the system, which means the time it takes to detect a certain motor-imagery induced neural state, as well as the state detection time, which means the time duration for which this neural state is recognized. Our investigations show that the neural patterns in the 0-8 Hz low-frequency band are essential for a short response time and that the patterns valid in the 8-30 Hz frequency band should be used to achieve a good state detection time. We show that both parameters, response time and state detection time, can be significantly improved when so-called cropped training method is used to train the deep neural net. Reaching a short response time and a good state detection time is significant for most continuous control applications. To the best of our knowledge, this is the first time the use of cropped training to optimize a continuous control application has been investigated.</p>

No abstract

<p>In this paper we report on our investigations on the use of a Deep Learning based brain-computer interface in the context of a continuous control application. A continuous control application is an application where a Deep neural network (DNN) is supposed to recognize defined motor-imagery related neural states from continuously measured EEG data stream and to initiate corresponding actions. Decisive for the quality of such an application is the achieved response time of the system, which means the time it takes to detect a certain motor-imagery induced neural state, as well as the state detection time, which means the time duration for which this neural state is recognized. Our investigations show that the neural patterns in the 0-8 Hz low-frequency band are essential for a short response time and that the patterns valid in the 8-30 Hz frequency band should be used to achieve a good state detection time. We show that both parameters, response time and state detection time, can be significantly improved when so-called cropped training method is used to train the deep neural net. Reaching a short response time and a good state detection time is significant for most continuous control applications. To the best of our knowledge, this is the first time the use of cropped training to optimize a continuous control application has been investigated.</p>