Probabilistic
filtering methods such as the Kalman filter and unscented Kalman filter enable
decoding of movement commands in brain-machine interfaces (BMIs).
In most prior work, the tuning models defining the relationship between desired
movements and neural recordings have been held fixed after initial model
fitting. However, several studies of neuronal variability have shown that the
modulation patterns of many neurons change over time, making filtering with
fixed models sub-optimal. Updating tuning models without access to the desired
movements (or assumed proxies for them) for re-fitting is a difficult machine
learning problem and has not been previously demonstrated for real neural data.
In this study, we show that a Bayesian tuning model update method built upon
the unscented Kalman filter can address this problem. Our method periodically
updates tuning model parameters using its own filter output as training data
("self-training"). Our periodic updates are computationally light
enough so that when amortized as a background computation, they can keep up
with changing tuning in real time. To improve the accuracy of the filter output
used to update the tuning model, our method smooths
filter output using the Rauch-Tung-Striebel smoother.
We keep track of the uncertainty in our tuning model parameters and perform
Bayesian updates. Our update method was tested on neuronal ensemble data
recorded in two monkeys that performed joystick reaching tasks. The monkeys
were chronically implanted with multielectrode arrays
in multiple cortical areas, including primary motor cortex, dorsal premotor cortex, and supplementary motor area. We used
recordings collected over 4-6 months to perform off-line reconstructions. Our
update method significantly improved filtering accuracy versus an analogous
filter with a fixed tuning model. The results suggest that our update method
can be used to improve BMI performance for neuronal populations with
time-varying properties.
[Poster PDF]
Selection of
appropriate decoder is key to the success of any
brain-machine interface (BMI). Ideally, the decoder should accurately represent
the relationship between neuronal modulations and behavioral parameters and
capture the specifics of the required behavioral output. To meet these goals in
a BMI that enacts reaching movements, we developed a decoding algorithm called
the n-th order unscented Kalman filter for BMIs. This decoder improves the common Kalman filter by
allowing estimation of desired hand movements under non-linear models of neural
tuning to hand movements. It employs a quadratic neural tuning model in which
the instantaneous firing rates of movement-tunned
neurons are modeled by the sum of a baseline firing rate, terms for the x- and
y-axis components of position and velocity, and terms for the velocity
magnitude and distance from the center position of the workspace. This
quadratic neural tuning model predicts instantaneous firing rates significantly
more accurately than a linear neural tuning model in position and velocity. The
algorithm also augments the Kalman filter state with a history of hand kinematic variables in the n-1 previous time steps,
allowing an n-th order auto-regressive hand movement
model. This n-th order AR model produces better
predictions of the future hand kinematic variables
before incorporating the neural activity input and allows the neural tuning
model to capture relationships between neural activity and hand movement at
different time offsets. The algorithm was tested in experiments in which two
rhesus macaques implanted with multielectrode arrays
in cortical motor areas performed reaching tasks first manually, then through a
BMI. We demonstrated superior performance of the new algorithm compared to
other commonly used decoders in both off-line reconstruction accuracy and
on-line, closed-loop BMI-controlled task performance.
[Poster PDF]
We have developed
a custom brain-machine interface (BMI) suite using Microsoft Visual C++ that
implements an array of BMI tasks. This BMI suite can be used for biological
inputs of single-unit and multiunit extracellular
recordings, local field potentials (LFPs), electromyograms (EMGs) and
limb-movement parameters. Extracellular recordings
are automatically spike-sorted using waveform templates generated by a modified
classification expectation maximization(CEM) algorithm and real-time sorting by
Multichannel Acquisition Processors (Plexon, Inc.). Template generation for two hundred units takes
approximately five minutes. Inputs can be selected via custom graphical user
interface (GUI), allowing arbitrary combinations to be selected for BMI
decoding. Decoding algorithms are modular and interact using a common interface
with the BMI software suite, allowing for easy addition of new decoding
algorithms. The output of the decoding algorithms is sent to the experiment
manager which handles stimuli, feedback, and experimental procedure logic. The
behavioral task, another modular component, can be switched in the experiment
manager without affecting the operation of the rest of the BMI suite because
the suite uses a standardized interface that smoothly links with all tasks. LAN
and internet communication via the UDP protocol allow the suite to send kinematic data to other BMI workstations or collaborators
around the globe in real-time. Kinematic variables
and algorithm predictions can be visualized in real-time using hardware
accelerated 3D graphics. All experimental data is recorded via a unified logging
system that automatically saves all pertinent variables every time they change.
Finally, the system incorporates a feedback loop based on spatiotemporal microstimulation through multielectrode
arrays implanted in brain sensory areas.
[Poster PDF]
Brain machine
interfaces (BMI) offer the possibility of restoring normal functions, such as
limb mobility, to persons with paralysis. Here, we report a novel algorithm for
decoding neural activity into commands for prostheses. This algorithm, n-th order Kalman filter, offers a better BMI performance
compared to previously used signal processing and machine learning methods. We
implemented this algorithm in a BMI that processed the activity of cortical
neuronal ensembles recorded in monkey (Macaca mulatta) engaged in a reaching task and converted the
neuronal signals into the predictions of the actuator movements. The algorithm
was optimized to utilize the abundant statistical information contained in the
movement profile and to detect stereotypical patterns of movements. The n-th order Kalman filter differs from the standard Kalman
filter by augmenting the state variables with a short history of the n-1 past
state variables. This allowed better prediction of the next state even before
incorporating the neural activity input. The comparison of the prediction
accuracy obtained with the n-th order Kalman filter
versus the standard Kalman filter and Weiner filter showed a significant
improvement in performance for several tracking and reaching tasks. This
improvement was achieved because the algorithm settled to a set of
stereotypical movement profiles instead of exploring an infinite number of
possible movements - the property of both normal voluntary movements and
practical neuroprosthetic applications.
[Poster PDF]
With the rapid
development of large scale, multi-unit recordings and their utilization in
brain machine interfaces (BMIs), algorithms for
sorting neuronal waveforms gained the utmost importance. We have developed
spike-sorting software for our BMI suit that records multichannel
neuronal inputs and performs real-time decoding of behavioral parameters from
these neural data. The quality of decoding hinges directly on the quality of
the sorted units extracted from neuronal ensemble activity. Utilizing a
classification expectation maximization(CEM) algorithm
written in C++, we can sort 128 channels of neuronal data in real time. MySql is utilized in order to distribute the computation
over multiple computers for optimal speed. In addition to the CEM, our
algorithm automatically filters the data to remove mechanical and electrical
artifacts. After the units are sorted, the corresponding sorted templates are
sent utilizing MySql back to main CPU. The template
waveforms are then written into Plexon settings and
utilized in Plexon's recording server to sort units
in real time. The templates can be updated by repeating this procedure during
the recording session. Our spike-sorting implementation is fast and flexible.
Utilizing 5 minutes of data to use for the template initiation, 128 channels
are sorted in approximately 10 minutes. These operations save more than 2 hours
typically required to sort such amount of multineuronal
data manually. Moreover, we have combined the sorting software with the
algorithm that selects the neurons which are best tuned to behavioral
parameters of interest. Using these neuronal subpopulations improves the
decoding by minimizing overfitting. The algorithm can
also adjust the clusters of sorted waveforms to improve the decoding accuracy.
[Poster PDF]
Our previous work
on the Traveling Salesman Problem and on the 15-puzzle suggested that humans
use pyramid representations involving hierarchical clustering to solve
problems. We tested this suggestion
by manipulating mental representation of the M+M puzzle, a sliding tile puzzle.
M+M puzzle can be presented to a subject in two, quite different ways. The results indicate that these two
presentations of the same problem were indeed treated as different problems by
the subjects. One presentation
allowed the subjects to solve the problem quickly, whereas the other made it
very difficult, if possible at all, to solve the problem. Simulation experiments showed that the
two presentations of the M+M puzzle lead to different pyramid representations,
which in turn lead to different solutions.
The presentation that was easy for the subjects was also “easy” for the
pyramid algorithm. These results
provide direct evidence for the role of mental representation in problem
solving.
[Poster PDF]
Last year we
presented results from experiments with human subjects, who were asked to solve
15-puzzle. We found that subjects
do not use distances among the states; instead they use “direction”. We conjectured that direction to the
goal is an emergent property of a pyramid representation of a problem. The pyramid representation is obtained
by hierarchical clustering in the problem space. The clustering does not assume the
existence of a global metric in the graph representing the problem, but only
existence of local neighborhoods on several levels of scale. This, in turn, assumes the existence of
an ordinal scale for distances.
Recently, we applied this model to 15-puzzle and demonstrated its
psychological plausibility. We also
showed the ability of the model to solve the classical problem of finding a
path around obstacles. Finally, we
applied this model to the M+M puzzle.
This puzzle has two different presentations (versions). Each version leads to a different
pyramid representation, and hence, to different solutions. We also tested human subjects on both
versions of this puzzle. The
results show a strong effect of the problem presentation on human
performance. This effect is similar
to that obtained in the simulation experiments with the pyramid algorithm.