J. Neural Eng. 17 (2020) 066009 W Li et al
To reduce the time required to collect new
samples, an efficient method is to take advantage
of the historical data to reduce the demand for
new samples. There may be three main solutions,
one way is to design an adaptive decoder which
can complete initial calibration by the historical
data and then self-recalibrate during the decoding
[16, 17]. Bishop et al developed a probabilistic self-
recalibrating classifier which could maintain stable
day-to-day performance without any daily supervised
retraining [17]. Jarosiewicz et al proposed a retro-
spective target inference-based decoder which could
achieve stable neural control quality for long periods
of self-paced iBMI use without the need for disruptive
recalibration tasks [16]. For the adaptive decoder, no
new current data need to be collected before iBMIs
using, however, the self-recalibration of the adapt-
ive decoder during the decoding will take some cal-
ibration time and increase computational burden.
Another way is using a more powerful decoder based
on deep neural network, which has good generaliza-
tion ability to deal with the nonstationarity. Sussillo
et al developed a new multiplicative recurrent neural
network iBMI decoder that successfully learned a
large variety of neural-to-kinematic mappings and
achieve more robust decoding performance for differ-
ent datasets [18]. This method needs a large amount
of historical data to finish the training of the decoder.
The last way is aligning the current data to the his-
torical data to diminish the nonstationarity. This way
tries to diminish the disparities between the historical
(source domain) and current data (target domain) in
order for the decoder calibrated by plenty of historical
data from the source domain to be generalized to the
target domain [19]. This is a typical domain adapta-
tion (DA) task [9, 10]. After aligning the current data
to the historical data, the decoder calibrated by the
aligned historical data can decode the aligned current
data properly without requiring additional recalibra-
tion. Therefore, we focused on this way in this study.
In the DA, the feature spaces of the source and
target domains are the same but have different mar-
ginal probability distributions. DA approaches can
handle the disparities between the source and tar-
get domains, and effectively apply knowledge learned
from the source to target domain [5, 20, 21]. DA
has been widely used in machine learning research,
such as computer vision [22, 23], natural language
processing [24, 25], bioinformatics [26–28]. In recent
years, DA has been applied to iBMIs, and some stud-
ies have been presented. Our previous study pro-
posed a principal component analysis (PCA)-based
domain adaptation (PDA) method to calibrate the
decoder [5]. Taking advantage of historical data from
the previous day, we used small current sample
set to calibrate the decoder and achieved favorable
performance. Dyer et al introduced a distribution-
alignment decoding method (DAD) which found a
low-dimensional mapping of current neural activities
to match the distribution of historical motor variables
by minimizing their Kullback–Leibler divergence
[29]. Lee et al proposed a hierarchical formulation
of optimal transport which leveraged clustered struc-
ture in data to improve alignment and could achieve
better performance than the DAD method [30].
Pandarinath et al proposed a latent factor analysis
via dynamical systems (LFADs) which used artifi-
cial recurrent neural networks to infer latent dynam-
ics from single-trial neural spiking data and could
accurately predict motor variables [31]. Similar to
the LFADs, Farshchian et al proposed an adversarial
domain adaptation network (ADAN) based on Gen-
erative Adversarial Network, which could align neural
recordings from different days [32]. Comparing with
two other methods (canonical correlation analysis
(CCA) and Kullback–Leibler divergence minimiza-
tion) proposed in the same study, the ADAN achieved
the best performance. The CCA was also used in
another study to extract stable low-dimensional latent
dynamics across different days [33]. Degenhart et al
proposed a manifold-based stabilizer to align estim-
ates of the manifold obtained from nonstationary
neural recordings from different days, which could
maintain iBMI performance in the presence of neural
recording instabilities [34].
These DA methods could effectively diminish the
disparities between the source and target domain
to maintain robust decoding performance. However,
most methods described above need to collect suf-
ficient target samples for finishing the alignment,
which will affect the real-time performance in online
decoding. Moreover, these methods either use only a
single-day historical dataset as the source domain or
combine the historical data from multi-days together
to be one source domain, they do not consider multi-
source domains by using each single-day historical
dataset separately as a source domain. Taking the PDA
method as an example, the PDA is a single source DA
method that only uses historical data from the pre-
vious day as the source domain. It works under the
assumption that data from the previous day is the
most similar to the target domain among all histor-
ical data [5]. However, this assumption is not always
satisfied because the nonstationarity of neural sig-
nals is unpredictable. If this assumption is not satis-
fied, using other source domains may achieve better
performance, which is also the case for other meth-
ods. Therefore, choosing the historical dataset that is
the most similar to the target domain or maximiz-
ing their use to achieve better and more robust per-
formance needs further study. Finally, most studies
described above focus on continuous decoding, such
as decoding continuous hand velocity and move-
ment trajectory from neural signals [29–34]. While
continuous decoding is very important, construct-
ing classifiers to decode discrete actions is also an
important part of the iBMIs [17, 35]. In the applic-
ation of prosthetic hand control with the iBMI, the
2
