FRIEDRICH, Rainer, Olfactory Pattern Classification, Mar 27, 2014
<p class="MsoNormal" >FRIEDRICH, Ranier, Olfactory Pattern Classification, Mar 27, 2014</p>
<p>
This is at Stanford in Clark Aud at Noon on Thursday; Neurosci Institute Seminar.
(This is a visiting superstar talk.)</p>
<p>
He is from Miescher Instit in Swtizerland. This is work using adult zebrafish.</p>
<p>
overview of sensory processing:
sensory input goes thru some transformation and recoding and then
goes to filtering/ Identification/ / memory/ concepts/ association.</p>
<p>
this is zebra fish work;
you can cut off the head and get brain and put it in a dish for several hours.</p>
<p>
representation of odor mixtures are NOT a superposition of component reps
Yaksi, von Sain et al Nat Neurosci 2009;</p>
<p>
olfac bulb goes to olfac cortex (and vice versa)</p>
<p>
to do synthetic object repres and pattern classification and auto-assoc memory...</p>
<p>
two critical computations in the olfac bulb:</p>
<p>
1) <b>pattern equalization</b> and
2) <b>pattern decorrelation</b>.</p>
<p>
Olfac Bulb sensors and neurons*</p>
<p>
an odor will activate particular configurations of <b>glomeruli (the sensors)</b> in the olfac bulb</p>
<p>
there are also interspersed <b>interneurons</b> and then the <b>Mitral cells
</b>are the output cells of the bulb.</p>
<p>
he uses 2 photon Ca imaging to see the output.</p>
<p>
So, first issue we will discuss is pattern classification:</p>
<p>
(But first we need to<b> normalize the intensities of firing</b>)
(he mentions perceptron invented by Frank Rosenblatt)</p>
<p>
he shows a perceptron like matrix</p>
<p>
input neurons 1 thru N... so, some combo of neurons gets activated for a particular odor
and then you have a summation neuron that is sensitive to that combo.</p>
<p>
for this to work: you must first have intensity normalization
(otherwise, one component will distort sum)</p>
<p>
the glomeruli are the input cells of the olfac bulb.</p>
<p>
to analyze this circuit <b>we targeted an inhibitory neuron</b>
called the "<b>short axon cells"</b> and</p>
<p>
modified those with GFP and GCaMP2 these are GABAergic. (inhib cells)</p>
<p>
when we incr concentration of an odorant we get more firing in these short axon cells
(cuz they get direct input from sensory afferents.</p>
<p>
but... what is their effect on mitral cells?...</p>
<p>
surprsiingly, they (the short axon interneurons) have a dual effect on mitral cells</p>
<p>
they did optogenetics to see the effect of these cells...</p>
<p>
post doc showed that <b>a gap junction current was important</b>.</p>
<p>
<b>gap junctions (btwn the inhibitory interneurons and the mitral cells)
are powerful enough to make the mitral cells fire!
(ie, they can over-ride the gabaergic effect</b>)....</p>
<p>
there is dense and widespread connectivity...</p>
<p>
his conclusion is that short axon cells transmit information about stimulus intensity.</p>
<p>
perhaps they are involved in equalizing intensities.</p>
<p>
you lose invariant representation if you turn off GJs = gap junctions with
GJ blockers.</p>
<p>
so, how do you equalize intensities?
one way of doing that is by <b>dynamic range compression</b>
(ie boost the weak stimuli and attenuate the strong stimuli)</p>
<p>
this is done in modern cameras eg....</p>
<p>
DRC = dynamic range compression...</p>
<p>
the gap junction can boost weak signal effect on mitral cells and
the GABA output is used to attenuate strong inputs.</p>
<p>
it also uses DRS = diversification of response sensitivities
(this is similar to HDR high dynamic range blending in cameras)</p>
<p>
this works to equalize the mean and the distribution.</p>
<p>
he shows cluster dagrams that show how this helps (by making tighter clusters)</p>
<p>
*</p>
<p>
now how does this work to mediate pattern decorrelation....?</p>
<p>
want to do this to increase pattern separation btwn 2 odors
(to get better autocorrelation eg)</p>
<p>
he shows firing rates in 58 distinct mitral cells for different odors.</p>
<p>
he shows firing rate over about 2 seconds...
they have transients (in their firing rates) but then stabilize.</p>
<p>
initially they are v similar (btwn odor 1 and odor 2) BUT...</p>
<p>
after a second or two.. the patterns separate and are quite distinct.</p>
<p>
see Friedrich and Laurent in Science in 2001...</p>
<p>
these patterns are highly corr. with animals ability to discrim odors.</p>
<p>
he shows repr in high D state space.</p>
<p>
the principal components diverge with difft odors (over a couple of seconds)</p>
<p>
he shows odor 1 = phenylalanine and odor 2 = tryptophan.</p>
<p>
how abt discrimianting George Clooney from (a guy who looks v similar to George)</p>
<p>
one way to probe is to morph George into the stranger and vice versa... video clip...</p>
<p>
does same thing with odors (morph these from 1 to 2)</p>
<p>
as you change concentrations... you get a sudden jump from pattern 1 to pattern 2.</p>
<p>
(this is evidence for discrete represenation to increase separation of patterns)</p>
<p>
see Niessing and Friedrich in Nature in 2010.</p>
<p>
these data are from 2 photon calcium imaging.</p>
<p>
what is the mechanism of decorrelation (btwn reps of odor 1 and odor 2)?</p>
<p>
how do you get contrast enhancement?</p>
<p>
(should be the opposite of above (equalization of intensity))
ie should make weak signal weaker and strong signals stronger.
that's the classic hypothesis.. but he does not believe it.</p>
<p>
experiment: dump dopamine into the bath and it enhances contrast...
cannot be explained by hypothesis...</p>
<p>
what does the data show?</p>
<p>
data shows that it is NOT by contrast enhancement...</p>
<p>
mechanism is MORE complicated...</p>
<p>
so, second hypothesis... (from a math postdoc in his lab)
involving neural networks...
this is by Martin Wiechert Nat Neurosci in 2010.</p>
<p>
theorem 1: thresholding alone invariably causes decorrelation
(ie, if you just raise the threshold that always increases separation) AND
theorem 2: recurrent connectivity amplifies this threshold induced decorrelation AND
theorem 3 this decorrelation is increased with sparse connectivity and high baseline activity...</p>
<p>
summary: olfac bulb does at least 2 operations:</p>
<p>
1) equalizing responses (normalizing amplitudes)
to create incr clustering and</p>
<p>
2) decorrelation (incr separation of repres of odors)</p>
<p>
seems important to understand higher levels (beyond the mitral cells)
(here he switches topics to block face electron microscopy EM)</p>
<p>
they use W. Denk serial block-face EM....</p>
<p>
takes images of the block face (not the slice that was cut off)</p>
<p>
uses this to generate connectivity patterns...</p>
<p>
this is still from Zebrafish</p>
<p>
voxel sizes are 3 by 3 by 25 nm !!!!</p>
<p>
this is from adult zebrafish... can also increase the size of the sections...</p>
<p>
beautiful EMs !</p>
<p>
have traced every single neuron ! (in the olfac bulb: all 900 !!!)</p>
<p>
only missing about 0.1% of dendrites...
can reconstruct every single neuron...
Questions for audience:</p>
<p>
Q1 how long does animal take to make distinction?</p>
<p>
(the steady state in the mitral cells is reached in 400 to 500 millisec)</p>
<p>
but the behavior is slow... like 19 seconds...</p>
<p>
but for a rat is perhaps 300 msec and for humans ... many seconds...</p>
<p>
Q2 can zebrafish disting mixtures of odors</p>
<p>
A mixtures are non-linear... one can mask the odor completely
or can be perceived differently than either component...</p>
<p>
Q (Fisher) are amino acids a natural stim for zebrafish?
A yes....</p>
<p>
Q can you silence short axon inhib cells in vivo (by optogenetics) to see the effect on behavior?
A we would like to... tech difficult.</p>
<p>
</p>
<p>
</p>