[ExI] another step towards uploading

[Anders Sandberg]

Ah, I think you misread me (in an interesting direction). I was not 
talking about dendritic tree computation, just finding out what was in 
the different synapses. If you do not know whether they are 
glutaminergic or dopaminergic, it doesn't matter how much you try to 
brute force what the rest of the cell is doing.

I think people tend to overestimate how hard it is to run a nonlinear 
neuron model. The big problem is getting the parameters right, not the 
actual simulation - that is mostly a matter of running a lot of HH-like 
equations, and optimizing for your available computational substrates. 
Getting enough data from tissue to pin down the simulation parameters 
*and* check that it produces sensible results (especially since the data 
collection might have ruined the cell for comparision and testing), that 
is the challenge. A super-resolution scan might still be worthless if it 
doesn't tell you what you need to know.


On 2013-10-04 05:11, Rafal Smigrodzki wrote:
> On Wed, Oct 2, 2013 at 5:49 AM, Anders Sandberg <anders at aleph.se> wrote:
>> On 2013-10-02 05:05, Rafal Smigrodzki wrote:
>>> A technique capable of directly reading nanoscale molecular
>>> distributions of multiple molecular species (a prerequisite for
>>> individual brain physiology reconstruction) is the holy grail of brain
>>> uploading. Near-field MRS could be it, perhaps.
>>>
>> I wonder how much and what information is actually needed to deduce synapse
>> types. Obviously, just cataloging the local chemical species would give a
>> ground truth. But we also know vesicle size and electron density gives about
>> a bit of information (roughly, excitatory or inhibitory). Voltammetry
>> reveals levels of dopamine, serotonin and noradrenaline. No doubt other
>> markers indicate other properties, likely in a noisy and overlapping manner.
>> But from a machine learning perspective, if we had good data sets of
>> multimodal data we could try training classifiers that could tease out a
>> proper decision tree. Comparing with real ground truth is of course
>> essential, so we need a "chemical dissection method" for synapses to use as
>> test set. But I suspect that in the end we will end up with surprisingly
>> simple statistical models.
> ### I think you are right about the simple statistical models
> substitutable for large chunks of dendritic trees but I would expect
> that the translation from a physical description of a neuron to a
> simplified equivalent might be very complex. There is a fair amount of
> analog computation going on in the dendritic trees, where the shape of
> the tree and the precise relative location of synapses modify the
> summation of inputs from synaptic strengths. After neural scan you
> would have to do some very heavy brute-force computations to yield the
> simplified models. All should be doable though, as long as the scan is
> of sufficient quality.



-- 
Dr Anders Sandberg
Future of Humanity Institute
Oxford Martin School
Oxford University