Page 165 - Handbook of Biomechatronics

162 Domen Novak

character with approaches such as the matrix-based speller (Section 2.5) and
the system shows the selected character to the user, then checks the EEG for
an ERP. If an ERP is detected, the character is either immediately deleted
(and the P300-based selection process is restarted) or replaced by the second
most probable character (Schmidt et al., 2012; Sp€uler et al., 2012). In able-
bodied participants, such error correction has been shown to increase writ-
ing speed by 40% compared to a P300 speller without error correction
(Schmidt et al., 2012); furthermore, improvements in writing speed can also
be observed in participants with severe impairments such as amyotrophic
lateral sclerosis (Sp€uler et al., 2012). Thus, these studies further validated
the potential of ERP-driven error correction in real-world BCIs. Other
recent studies have extended this approach to other realistic BCI applica-
tions, such as controlling humanoid robots (Salazar-Gomez et al., 2017).
In the long term, ERP-driven error correction is likely to become common
in a broad range of BCIs, as it does not require any additional hardware (it is
based on the EEG) and can significantly improve BCI performance.

2.9.2 Error-Driven Learning
The second possible application of ERPs is to perform error-driven learning,
where the underlying algorithms of the BCI are adapted in response to errors
(Chavarriaga et al., 2014). For example, Artusi et al. (2011) initially trained a
BCI classifier for recognition of fast vs slow motor imagery on a set of pre-
viously recorded EEG data. This dataset was then kept in the BCI’s memory.
When a user interacted with the BCI, incoming EEG was classified as fast or
slow motor imagery, and the result was presented to the user on a screen. If
no ERP was detected, the classification was considered correct, and the
newly recorded EEG signal was added to the dataset in memory together
with the classification result. At regular intervals, the motor imagery classifier
was then retrained using both the original EEG data and the data obtained
from the current user, gradually tailoring the BCI to the current user and
increasing its accuracy.
Besides retraining the BCI pattern-recognition algorithms, ERPs can
also be used to adapt the behavior of other machines. The user monitors
actions taken by an intelligent device; when the device performs the wrong
action (e.g., a mobile robot takes the wrong path or a humanoid robot makes
the wrong gesture in response to the user), an ERP is detected and the
device’s control algorithms are automatically updated to reduce the proba-
bility of that action being taken in similar future circumstances. A few prom-
ising studies in this area have demonstrated that humans generate ERPs in