[ExI] D-Wave's Quantum Computer

[Rafal Smigrodzki]

On Fri, May 17, 2013 at 1:25 PM, John Clark <johnkclark at gmail.com> wrote:

>
> 3) QUBO problem, Quadratic Unconstrained Binary Optimization problem, it
> concerns the minimization of quadratic polynomials. QUBO is a NP hard
> problem and is a pattern matching and image recognition technique used in
> machine learning. The enormously important protein folding problem is also
> QUBO as is anything where you have a lot of different things that can
> attract or repel each other and you want to arrange things in such a way
> that the entire system has the lowest possible energy.

### Patterns of input and output neural spikes can be observed using
cortical microelectrodes. Arrays of such electrodes collect data from
neighboring cortical columns, and each one detects only a vanishingly
small subset of the neural traffic. Obviously, the patterns must be
correlated to some extent, since the neighboring columns exchange data
all the time. One of the features of biological neural networks is
minimization of path length and the use of various mechanisms (e.g.
synaptic pruning) to control the amount of traffic resulting from a
given input.

It does seem to me there is at least a faint analogy between this
process and the minimization of potential energy inherent in protein
folding. Does it mean that D-wave chip could be used to adiabatically
calculate the lowest-complexity pattern of synapses capable of
generating the neural traffic observed by a microelectrode array?

Think about it this way - a protein sequence puts constraints on the
3D structures that can be realized from it, and the key is to find the
3D structure (or a set of them) that are most stable (i.e. have the
lowest potential energy). A microelectrode pattern puts constraints on
the networks capable of generating the pattern, and the key is to find
the simplest network reliably tracking with the microelectrode pattern
over time. In the first problem you move a molecular bond by a tiny
bit and calculate the potential energy impact on the rest of the
structure. In the second problem you change a single synaptic strength
(which could be 0, i.e. creating or removing a synapse) and see if the
resulting network better matches a time-ordered series of spikes in
the microelectrode recording. In both cases you use a bit of noise to
get you over local minima and makes the simulation more realistic.

How would an algorithm designer call the computation of network
structure from network traffic? Is it similar to QUBO?

If yes, then adiabatic computing might greatly reduce the amount of
data needed to upload a person.

Rafal