PILLOW, Jonathan, Parietal Cortex Dynamics in Decision Making, Princeton, May 16, 2016
<p class="MsoNormal" >PILLOW, Jonathan, Parietal Cortex Dynamics in Decision Making, Princeton, May 16, 2016</p>
<p>
Princeton Neuroscience Instit;
Dept of Psychology and
Center for Statistics and Machine Learning.</p>
<p>
this is at 5:10PM in MBC grad training series at Math Corner (Stfd on Mondays.)
big crowd today!</p>
<p>
title Unlocking single-trial dynamics in parietal cortex during decision making...</p>
<p>
motivation:</p>
<p>
the problem...</p>
<p>
sensory motor decision making (a la Newsome lab)</p>
<p>
sensory... coherent random dot motions</p>
<p>
LIP (in parietal cortex)</p>
<p>
long running collab with EJ Chichilnisky...</p>
<p>
will focus on latent dynamics (not just on the explicit repr)</p>
<p>
so, what do we mean by latent dynamics?</p>
<p>
previous... "direct" sensory encoding of model...</p>
<p>
we build a GLM = generalized linear model</p>
<p>
function of ( dot s)</p>
<p>
dot product... of responses given the inputs.</p>
<p>
in contrast to that...</p>
<p>
noise in generating X from S and also noise generating r (the response from X.</p>
<p>
probability of response given stim =</p>
<p>
integral of prob (response given x) andprob of x given s...</p>
<p>
bayesian model...</p>
<p>
with markov piece over time...</p>
<p>
the task...</p>
<p>
(a la newsome)...</p>
<p>
random dots decision making task (a la mike shadlen and bill n)</p>
<p>
lateral intraparietal area = LIP</p>
<p>
when animal plans to make a saccade...</p>
<p>
monkey is fixating in center...and there is an off center RF = response field...</p>
<p>
region where an LIP neuron responds to a saccade target...</p>
<p>
that represents the target...</p>
<p>
then motion of dots...</p>
<p>
the monkey reports his decision by saccading to the R or to L...</p>
<p>
can either wait for GO cue or do it spontaneously...</p>
<p>
shows video of moving dots...</p>
<p>
monkeys are great at this...</p>
<p>
control fraction of dots (+50% means dots moving to the RIGHT (mixed with 50% noise)</p>
<p>
animal integrates the evidence (drift diffusion model)</p>
<p>
acts when it crosses threshold...</p>
<p>
but also neurons that report out the animals decision...</p>
<p>
area MT...</p>
<p>
notion is that LIP is where the accumulation of evidence is occurring...</p>
<p>
now... does LIP look like gradual accum of evidnece or is it a step function?...</p>
<p>
BW.. Q is it indiv neurons in LIP or the whole region...?</p>
<p>
A: downsteam neuron reads out a population of neurons (but really talking abt many neurons)</p>
<p>
(many other things also repr in LIP)</p>
<p>
Shadlen and Newsome in 2001..</p>
<p>
see a ramping response where time t zero is when motion begins...</p>
<p>
(looks like evid accumulation)</p>
<p>
but... how abt dynamcis on single trials? (above was averaged across trials)</p>
<p>
two possibilities...</p>
<p>
1) noisy ramping (a noisy accum to a threhold...</p>
<p>
2) (as opposed to a discrete stepping (in many neurons) that on average look like evid accumulation.</p>
<p>
BW interested in noise (what is its nature)</p>
<p>
JP = Jon Pillow</p>
<p>
student comment... is this just difft neurons hitting a threshold at difft times.</p>
<p>
JP... let's defer that...</p>
<p>
JP all models are wrong! (all are toy models)</p>
<p>
but which toy model is best...</p>
<p>
JP 5 basic stim types: strong R, weak R neutral, weak L and strong L...</p>
<p>
*</p>
<p>
formalizing the models..</p>
<p>
formal bayesian approach...</p>
<p>
ramping ("diffusion to bound") model...</p>
<p>
say each neuron starts at some spike rate and then has slope upward beta sub c.</p>
<p>
add in noise at sigma 2.</p>
<p>
model is</p>
<p>
latent state is x sub t = x at t-1 + beta + noise...
so spiking response(t) given x at time t is approx poisson.</p>
<p>
(with each bin flip a coin (at to whether it spikes or not))</p>
<p>
this is continuous time (but we did 10 msec bins (to bin the raw data)</p>
<p>
in contrast</p>
<p>
the stepping model (discrete switchn)</p>
<p>
semi-Markov...</p>
<p>
NegBino...</p>
<p>
prob of stepping up is pi sub c</p>
<p>
this is a contin gamma distrib.</p>
<p>
now, simulate a few trials...</p>
<p>
hard to just solve visually (can't do it)</p>
<p>
what is the actual PSTH of the neuron...</p>
<p>
step time distrib..</p>
<p>
(model cannot do bursting etc.... need more detailed model)</p>
<p>
a critique... is the spiking really Poisson... (important assumption)</p>
<p>
seems robust to deviations...</p>
<p>
we at 200 msec after start of motion (but we also look at 150 msec etc)</p>
<p>
its robust whether earlier or later...</p>
<p>
we did not model under a 150 msec latency...</p>
<p>
eg we say... thisneuron starts accum at 200 msec (or has step up (as in model 2)</p>
<p>
all is done , modeling indiv neurons (should be a subscript i, says a student)</p>
<p>
(he addresses questions abt other model variations)</p>
<p>
Conclusion of this paper:</p>
<p>
all we get to see just the coherence of the dosta dn the response...</p>
<p>
we really want the param set theta that best fits...</p>
<p>
we use Markov Chain Monte Carlo (Gibbs sampling)</p>
<p>
a procedure for smapling joint distrib over latent variables...</p>
<p>
goes back and forth btwn model and assumed params.</p>
<p>
his student did 40X speed up using GPU based implementation.</p>
<p>
(Kenneth did this as a course project!)</p>
<p>
shows strong "out" motion (to the left</p>
<p>
0% coherenece...</p>
<p>
versus strong Rward motion...</p>
<p>
(like sampling the Volkswagen exhause and tryng to infer what's going on under the hood!)</p>
<p>
now, the main result...</p>
<p>
Latimer et al 2015...</p>
<p>
DIC deviant info criterion...</p>
<p>
penalty for incr # of params...</p>
<p>
31 out of 40 neurons have better fit by stepping model</p>
<p>
we have 25 (versus 6 neurons) with storn evidnece...</p>
<p>
so it seems that STEPPING MODEL is correct !!!</p>
<p>
Surya... suggestion... (but Surya has anticipate his next slide)</p>
<p>
his student Kenneth is a die-hard DIC enthusiast... (altho had similar esults with predicitve log-likelihood)</p>
<p>
we did simulation...</p>
<p>
great agreement with stepping simulaitons (ramping model was worse)</p>
<p>
Q how abt a few widely out of bound neurons? A could be...</p>
<p>
he argues... hugely more likely that the steppingmodel is correct.</p>
<p>
but there could be all kinds of heterogeneity.</p>
<p>
they did two animals...</p>
<p>
PSTH in spikes per sec.</p>
<p>
(he shows a slide that positively addresses Surya's comment)</p>
<p>
does the step just repr the animal's decision...</p>
<p>
seems not.</p>
<p>
population carries an accum of evidence (but each indiv is not correl with the ultimate choice)</p>
<p>
future direcitons:</p>
<p>
more detailed model...</p>
<p>
eg time-varying dynamics...</p>
<p>
SUMMARY</p>
<p>
formal comparison of rrampng vs stepping</p>
<p>
best model is of "steppers"</p>
<p>
Kenneth Latimer was the amazing grad student who did the math models</p>
<p>
and Alex Huk... long time collab.</p>
<p>
questions </p>
<p>
Q so, is accum occuring in indiv neurons at all?</p>
<p>
A how to do science...</p>
<p>
what we do (in his lab) is getting best model of our data...</p>
<p>
critique... not hypo driven (say his critics)</p>
<p>
Q talking abt mechanism...</p>
<p>
BW (reflectingon EJ's comment): </p>
<p>
JP got a lot of heat for denying the ramp model...</p>
<p>
BW... so, how big the effect size (also asked by Surya)</p>
<p>
BW aren't you saying that neurons have binary states... switching form state A to state B...</p>
<p>
any other examples of that?</p>
<p>
</p>
<p>
</p>