II. MARCO TEÓRICO
2.1 Marco conceptual
2.1.3 Capital intelectual organizacional
As it was previously mentioned, BCIs have been traditionally developed as a form of assistive technology for people with severe disabilities. However, the advances in EEG technology and the development of relatively inexpensive headsets made possible the rise of BCI systems that can be used by able-bodied users as an extension of their capabilities [Cecotti & Rivet, 2014;Wang & Jung,2011; Yuan et al., 2012], i.e., as a new input channel.
A field that has arisen from these advances is that of collaborative BCIs, in which several users aim at jointly controlling one device simultaneously. In order to achieve this, the EEG signals from a group of users are merged or decoded together, so the final command that is sent to the device is derived from their collective intentions, rather than from a single user. It is worth noting that improvements derived from this field can in turn open the door to new clinical applications for single-user BCIs [Finke et al., 2009]. For example, the current need for BCI systems to train a classifier at the beginning of each session in order to adapt to the user reduces the remaining time of practical use of the BCI [Fazli
et al., 2009a,b;Krauledat et al., 2008; Mora et al., 2015].
Model calibration (including feature selection and classifier optimisation and training) can be very time-consuming depending on factors such as the number of electrodes being recorded, so it is desirable for possible day-to-day use of this technology to reduce training times and increase transference, i.e., reduce the amount of data collected on the day to update a previously trained model (or classifier) while keeping a reasonable performance [Krauledat et al.,2008; Manor & Geva,2015;Marathe et al.,2016]. Collaborative BCIs are likely to work toward decreasing training times and develop new methods for cross-session and cross- user transfer [Huang et al., 2011; Sajda et al., 2010] for example through the creation of user databases, so that the EEG signature of a new user can be compared to those from the database for a “plug and play” experience.
The aggregation of signals in collaborative BCIs can be done in different ways [Cecotti & Rivet, 2014; Li & Nam, 2016; Nijholt, 2015; Poli et al., 2013a;
Wang & Jung, 2011], as shown in Figure 2.1.
The simplest method consists of performing averages of the raw EEG signals across several users prior to their classification (signal fusion level). In this way, a unique classifier is used for all subjects and there is a reduction in the inherent levels of noise in the signals [Cecotti & Rivet, 2014; Cecotti et al., 2014b; Jiang et al., 2015; Kapeller et al., 2014; Korczowski et al., 2015; Poli et al., 2013a]. However, this method may not be the most accurate, since the latency of time- locked ERPs (and, in particular, P300 components) varies both inter- and intra- subject depending on factors such as the level of attention, as was discussed in Section 2.2.1.
Figure 2.1: Different strategies to merge the brain activity from multiple users in collaborative BCIs.
this scenario, features are extracted from each user’s EEG. The fusion of features can be done by simple concatenation of them to form a unique feature vector for classifier training or any other combination [Eckstein et al.,2012; Wang & Jung,
2011], so only one classifier is used (as in the signal level approach).
Finally, the information fusion can be done at the decision level, in which the EEG data from each participant is used to tailor one classifier specifically for him/her. In this case, a decision merging step needs to be implemented. At this level, we should emphasise the work from Cecotti & Rivet [2014]; Cecotti et al. [2014b], who studied different modes of combining the BCI decisions on a P300-based collaborative BCI and a SSVEP multi-brain BCI. Their strategies for merging the classifiers outputs included majority voting, averaging classifiers’ outputs, and maximum and minimum values. They found that averaging the classifiers’ outputs provided the best performance.
A considerable amount of work has been conducted to establish which level of fusion is optimal, obtaining consisting results across laboratories and applications. In particular, the two approaches that are often compared are the single-trial averages across participants (i.e., signal level) and fusion at the decision level (usually averaging classifiers’ outputs to send a command). Since most of this work has been done based on different ERPs, given the inter-subject differences in latencies and amplitudes, it is not surprising that the best performance is obtained when information is merged at the decision level [Cecotti & Rivet,2014;
Cecotti et al., 2014b; Wang & Jung,2011].
Even though it is, in theory, possible to repeat trials for every participant also in the collaborative paradigm, it is expected that the classification will be done in single trials. Given the relatively big amplitude of the P300, this should not be a major problem for performance (especially if averaging across a sufficient number of users), but it could be a limiting factor for the detection of other ERP components, such as the N2pc.
Applications of collaborative BCIs include the simultaneous joint control of a single device. Wang & Jung[2011] showed that multiple users are able to generate and send a movement command to a prosthetic limb faster than single users. They also studied several methods of combining EEG recordings and reported an increase in the classification accuracy when the number of subjects was increased (they tested up to 20 individuals). Cecotti & Rivet[2014] obtained similar results on a P300-based matrix speller BCI in which up to 10 individuals were combined.
Poli et al. [2013a] simulated an offline cBCI by combining signals from pairs of users to control a pointer on a screen by means of their BCI mouse [Citi et al.,
their participants.
Collaborative BCIs can also be found in the field of group decision making. In a scenario where it is not possible to average epochs over several trials (e.g., a person cannot be asked to make the same decision multiple times), aggregating the signals from a group of users in order to achieve a better outcome while re- ducing the level of noise has proven to be a useful technique [Poli et al., 2013b]. Studies in this area of research tackle two problems: (1) whether group perfor- mance for decision making beats that of a single person, and (2) whether there is an optimal group size. Poli et al.[2013b] used a collaborative BCI to integrate the brain activity from up to seven subjects in a decision making task based on visual perception. By increasing the group size, errors were reduced. In further work, they extended their research (improving their results) by including non-EEG (i.e., behavioural) features, e.g., response times [Poli et al.,2014].
Lastly, collaborative BCIs have been applied to visual search, using both search arrays [Valeriani et al., 2015b] and naturalistic images [Valeriani et al.,
2015a]. The authors of this research found improvements in group performance on a target detection task when using a confidence estimator based on EEG signals and response times (hence using a hybrid BCI) to weigh each participant’s response on a trial-by-trial basis.
Collaborative BCIs can also be used as passive systems. In fact, the origin of cBCIs can be tracked down to 1965 if this type of systems is taken into account. Back then, they were known as hyperscanning systems [Babiloni & Astolfi,2014]. Hyperscanning allowed researchers to discover that collaborative and competi- tive tasks have different effects on the connections in the brains of participants performing behavioural experiments (e.g., Astolfi et al. [2010]).
The concept of hyperscanning recently surfaced as passive multi-mind BCIs, examples of which can be found in the work of Hasson et al. [2004] and Hasson et al. [2008], in which the authors assessed the effect of feature films on brain activity during free movie watching. The main result of their work was to show that aspects such as movie content, editing and directing style have a direct im- pact on the level of control over the viewer’s brain activity. Later studies used passive multi-mind BCIs to show the high level of inter-subject correlation during natural vision [Bridwell et al., 2015]. This discovery makes it possible to study the brain’s naive responses to stimuli by averaging signals across multiple users, hence increasing the low signal-to-noise ratio that is typical in EEG-based BCIs. Moreover, the high time resolution provided by EEG systems allows researchers to use this technique in multi-brain BCIs, for example, in those based on ERPs, which traditionally rely on multiple repetitions of a stimulus in single-user inter- faces [Jiang et al., 2015;Kapeller et al., 2014; Korczowski et al., 2015].