KAO, Jonathan, Brain Machine Interface BMI, July 27, 2016
<p class="MsoNormal" >KAO, Jonathan, Brain Machine Interface BMI, July 27, 2016</p>
<p>
intro by Krishna Shenoy, his advisor.</p>
<p>
This is at 2:15PM in Clark S360... a thesis defense.</p>
<p>
OUTLINE</p>
<p>
the target patient population: paralyzed patients like</p>
<p>
Christopher Reeves (1995: accident , c-spine injury). died in 2004.</p>
<p>
BMI: tries to overcome motor neural paralysis</p>
<p>
implants Utah Array with 100 trodes. (implants two in their monkeys: one in M1, the other in the dorsal premotor cortex)</p>
<p>
shows 96 channels fro Utah Array. spikes from nearby neurons.</p>
<p>
spiking neuron signals have the highest signal for control. .</p>
<p>
S uses visual input (feedback) to correct their movement of the computer cursor.</p>
<p>
BMI interprets the brain signal to move cursor.</p>
<p>
shows a dozen references from other labs.. many labs involved..</p>
<p>
shows a video of woman with ALS.</p>
<p>
free, paced typing using the OPTI II keyboard.</p>
<p>
moves cursor just by thinking about it.</p>
<p>
she can type at 8 words per minute. she must be able to click the mouse when she reaches the target.</p>
<p>
Algo is ReFIT-KF... etc.</p>
<p>
we need faster and more accurate cursor control</p>
<p>
must have accuracy eg for click signal..</p>
<p>
robustness... system may change over sessions</p>
<p>
must be robust over years and maybe deacdes.</p>
<p>
we worked with macaques...</p>
<p>
homologous to humans...</p>
<p>
monkeys enjoy these... will do this > 3 to 8 hours per day.. they like it.</p>
<p>
Xk = movement at time k to position X.</p>
<p>
shows 4 neurons... one per channel.</p>
<p>
draws a window that is 20 ms long... and just count spikes</p>
<p>
that is Y at time k.</p>
<p>
decoding every 20 ms.</p>
<p>
find a funciton f that maps from yk, y at k-1, etc) into x sub k.</p>
<p>
monkey is f fasst.</p>
<p>
faster than the human!</p>
<p>
bitrate is 3.8 bps... multiply by 2.5 to get 10 words per minute.</p>
<p>
monkeys will not arie.. he jokes.</p>
<p>
he holds over it for 0.5 seconds to select it (not giving a click signal to enter it)</p>
<p>
this is in JNE in 2014.</p>
<p>
Hellow World My name is Jenkins and...</p>
<p>
but now we have click signal...</p>
<p>
he shows a 2 d grid for neuron 1 on X axis and neuon 2 on Y axis.</p>
<p>
shows how to decode a CLICK signal (ie enter that letter)</p>
<p>
LDA = linear discriminant analysis...</p>
<p>
success rate is 70.0% but this is NOT better than just holding for 1/2 second.</p>
<p>
why? cuz state space overlaps</p>
<p>
we should really be decoding the MOVE state most often.. MOVE means moving from one letter to the next.</p>
<p>
perhaps 95% of the time.</p>
<p>
uses HMM hidden Markov Model...</p>
<p>
shows state transitions for HMM...</p>
<p>
if in the MOVE state then stay in that until you're sure that you're in the STOP state, then stay there</p>
<p>
achieves a 50% incr. in speed!!!</p>
<p>
now we have even more complex transitioon models</p>
<p>
cuz start in slow state, then move to fast move state. etc. then when close to the target.</p>
<p>
so, we can get SOTA control with HMMs...</p>
<p>
compares monkey J and monkey R.</p>
<p>
same with robotic arm...</p>
<p>
*</p>
<p>
neural data is y sub k
kinematic data is the x sub k.</p>
<p>
the representational vie is that Yk = f(Xk)</p>
<p>
PD = preferred direction of reach angle.</p>
<p>
this is called the velocity Kalman filter.</p>
<p>
if moving to the R now the probab. will STILL be at time k+1.</p>
<p>
but these sim[le tuing curivng for reaches are not static.. they depend on the speed of typing, etc.</p>
<p>
but causality is backwards... (we really want it flipped aorund)</p>
<p>
he shows PCA principles.</p>
<p>
eg to reduce 3d curve down to 2d for an S cruve.</p>
<p>
he shows how noisy the nueral data is.</p>
<p>
shwos path of cannonball as a parabola.</p>
<p>
say, your video of ball is low resoltuion.</p>
<p>
shows cannonball going off screen.</p>
<p>
cannonball canno tgo down and then UP in mid-flight... we know Newtonian mechanics.</p>
<p>
study by Churchland et al.</p>
<p>
shows neural data evolving in lawful way as CCW rotation.</p>
<p>
from Mark Churchland.</p>
<p>
so, we change the model to a dynamical model of the neural data.</p>
<p>
arrows describe the flow field...</p>
<p>
this corrects the prior limits...</p>
<p>
evovles from Sk-1 to Sk to Sk+1 and those predict both theneural data and the positional data.</p>
<p>
our new appraoch gets up up to 3.8 bits per sec.</p>
<p>
last part of talk...</p>
<p>
robustness (stability)</p>
<p>
trode arrays may (biofoul) and lose signal.</p>
<p>
shows great data one month after implantation...</p>
<p>
but two months after that...</p>
<p>
we see fewer signals than before.</p>
<p>
eg now have fewer neurons that we record.</p>
<p>
nos, say, fast forward by a year and see only a tiny # of neurons.</p>
<p>
but even tho some neurons are lost.. we wtill have that data.</p>
<p>
shows terrible performance even they artif dtop 110 neurons from 200.</p>
<p>
but now use "hysteresis" algo to rescue the decoder and
hence to prolong life of BMI device.</p>
<p>
want to develop dynam system with all the neurons and then keep that
when neurons drop out...</p>
<p>
ie remember the neural dynamics from early, good data.. and then always use those
(like learning the dyanmics of the cannonball)</p>
<p>
get a huge incr in perforamnce to incr robustness.</p>
<p>
our new algo incorproate prior K. across the lifetime of BMI.</p>
<p>
QA</p>
<p>
</p>