[coulls]

This thread is specifically targetted toward Rob and DavidOlmsted, as they both have infinitely more knowledge about the particular area of FPGA processors than I do, but feedback is welcome from all members of the community.

After reading the 'Electronic Evolution' post in the 'Model and Simulation Topics' forum, and I was struck by something. At one point, FPGAs are mentioned as a possible vehicle of massively parallel computing needed for a hardware version of the hierarchical system mentioned in 'On Intelligence'. DavidOlmsted is quick to point out, however, that the FPGAs operate on binary logic and each have discrete-time sequential processing, neither of which lends itself to the task of 'brain simulation' very well. I went back to my thread on the possible use of quantum properties in the cortical algorithm, which I agree doesn't seem likely upon closer inspection. However, what if we were to use these quantum effects in a brain chip. Granted, this would not be a faithful macro-scale simulation of neuro-transmitter emission and reception as the brain really works, but quantum properties seem to lend themselves very easily to the problems we have in creating hardware to properly simulate the brain.

First, quantum properties by nature operate on a type of 'fuzzy' logic. This eliminates DavidOlmsted's problem with FPGA processors, as the use of quantum mechanics allows for various logic operations to be performed on 'qubits', as they are called, which modify the probability function for certains states in the system. We can get these to behave essentially exacty like a binary logic system, but by nature they are very much based on probablistic logic. Second, quantum mechanics also lends itself very nicely to massively parallel computation. Here we can have many quantum states which represent different solutions to a particular problem, all of which can be operated on simultaneously by a single operation, and the correct solution automagically appears when the superposition collapses (assuming your use of operators on the system is correct).

It seems pretty clear to me that quantum computing properties offer solutions to two of the big problems with traditional processors, namely parallel computation and probablistic logic. Unfortunately, if I remember correctly the largest quantum computer built to-date has something like 7 qubits...certainly not enough to produce anything but the most trivial operations.

What does everyone think? Does it have promise? Someone have a working quantum computer hanging around that we can try out?

[Arne Gleason]

Quantum computing is one of those things where I think I get the concrete examples (with strenuous effort) but utterly fail at being able to ponder it abstractly. I see how bits in superposition work wonderfully for factorials but I can’t imagine how they would benefit the majority of computing problems (I guess I’m very limited – If you showed me how to use a transistor to build an amplifier, that’s probably the only use I could imagine).

All that being a general disclaimer, I do anticipate a confluence between quantum computing and engineered intelligence. Not because one needs the other -- just because computing things faster allows for approaches that aren’t economical at slower speeds (or maybe quantum computing will be only useful for the narrowest of applications, and speed isn’t the problem anyway).

[DavidOlmsted]

Very perceptive Coullis. Asynchronous information processing (as done in the brain) is mathematically very close to quantum mechanics. The only difference seems to be that asynchronous information processing requires defined pathways (neurons) where quantum mechanics occurs in unconstrained "free space".

As you noticed the electrical potentials within neurons are fuzzy-like being wave-like and declining exponential-like just like quantum potentials at a much smaller scale. A circuit composed of neuron-like operations directly connected (not using action potentials to communicate) would be like a quantum circuit. Or in other words a circuit composed of directly connected quantum wells would be like a neural circuit.

However, this information processing would not be like a finite state machine with absolutely defined states. Instead the best that could be done at any given time would to pick the most probable state in a competitive process. This means that the circuitry must be able to handle uncertainty as in determining how much uncertainty is acceptable for any given problem (degree of mathematical precision).

Right now all quanum wells are perpendicular to a chip's surface since they are made epitaxially (layers of chemical deposits). One cannot make directly connected quantum level circuits this way. But if someone ever figures out how to make and connect quantum wells in a circuit then we would have a new "neural" way of doing information processing.
_________________
Click to go to my site at neurocomputing.org

[coulls]

Thanks for the reply David (I hope it is alright to refer to you as David, rather than your nickname). One of the points of my post was that quantum computing can be treated like a finite state machine, however it can also be treated differently, at least as far as my understanding of quantum mechanics goes (and that is not very far, let me assure you).

I envision the following:

Picture the various cortical regions as sets of quantum superpositions, where each column, for instance, would be an entangled superposition. These superposition would be entangled with the columns at the higher cortical region, such that changes in the probability function for the higher cortical regions' columns have an effect on the probability function for the lower cortical columns as well. Now, we have (approximately) the quantum equivalent of the cortical hierarchy, minus the specifics like that thalamus. I know that the thalamus is probably important in some respect, but my goal is to draw out a deeper discussion of the implications of quantum mechanics in the implementation of this algorithm.

Anyway, we have this hierarchy of entangled states. Now, some input comes along and alters the probabilities of the lower level column states, thereby influencing the upper level probabilities as well. The beauty, as I see it, is that this captures the fact that changes at the higher level are very gradual over time. For instance, after several inputs, the quantum equivalent of the hippocampus may have a probability that has finally converged to 1 for some state, now upon continued input, that value changes to reflect the probable 'high level' states for the input seen thus far in the superposition. So, in fact, getting a proper classification simply (yea, simply my foot) comes down to observing the state and collapsing the superposition. This will then probabilistically select a state for each column, and in turn for the entire hierarchy. Again, this is the rough and dirty idea...and may be way out in left field in reality, but it may be an interesting concept to explore - especially from someone who is more well versed in quantum mechanics than myself.

Perhaps I should think the whole thing over more before continuing to post, but I enjoy the sounding board (no pun intended) here at the 'On Intelligence' forums.

Also, as a note to your quantum circuit comment, I believe that quantum tunneling allows the 'transportation', so to speak, of electrons between quantum wells with a certain amount of energy required to cause the tunneling to occur. Again, it has been a while since I've even lifted a book about this stuff, but tha was just my recollection.

[DavidOlmsted]

David is fine.

Coulls, you are reaching for the asynchronous information processing concepts not found in finite state machine theory but only in quantum mechanics and the brain. Quantum mechanics has no external univeral clock to define a sequence of states, instead the states can only be determined when a decision is forced (the system observed). The lowest level of forced decision making in the brain is the generation of action potentials themselves via a threshold test. Higher levels of forces decisions are based upon comparing two or more action potential inputs.

So neurally, the equivelent of quantum superpositions are the superpositions of the analog membrane potential generated by each synapse. A burst of action potentials into a synapse produces a increasing or decreasing exponential function (like the quantum potential outside of a quantum well). Wave effects are only produced by a train of bursts. This is where neural processing is different from quantum effects. In quantum mechanics waves always occur when the energy of the system is greater than the boundry conditions. Waves in neural systems have to be generated.
_________________
Click to go to my site at neurocomputing.org

[Rob]

Research has been done in clockless FPGA design. Philips Electronics, I think, developed an entire pager without a clock because of signal interference issues. Right now they have the only tool for analyzing asynchronous digital logic. There is also a company called Soneticom that was experimenting with FPGAs that reconfigured themselves on the fly. But I don't know where that research went.

Ultimately, neurons make a binary decision - fire or not fire. I don't see why digital logic can't represent that. Signal firing rate could be easily adjusted, and signal strength could be represented by a several bits, if need be.

[Arne Gleason]

Hi Rob,

[Rob]

Arne,
I'm not sure if that conjecture is true. All I'm saying is that neurons basically sum their inputs and decided whether to fire or not. It doesn't matter what state the rest of the system is in, only what the inputs to a give neuron are doing.

Let's say we have a neuron 'A'. This neuron has 100 inputs. 'A' has a summation function that says "once my threshold of 70 is crossed, I will fire." That threshold of 70 (I just picked 70 out of the air) can be reached in many different ways. The input neurons will convey their intensity by how quickly they fire. Neuron 'A' may make a firing decision every 1ms. So it basically collects inputs for 1 ms, then decides to fire. If 90 of the 100 input neurons send a signal, that may cause the threshold value to be reached. Alternatively, if 5 neurons go crazy and fire 20 times each during that 1ms, that too may set the threshold above 70. There are some issues to be addressed, but I believe all that can be represented in digital logic.

The biggest problem with a digital implementation is the vast majority of connections that need to be made. Anyone who had done digital design, particularly with FPGAs or CPLDs will tell you that you usually run out of routing resources before you run out of logic gates. But Hawkins made a good point in his book. Because of the slow rate at which neurons fire, and the fast rate at which hardware runs, the connections can work the telephone system. Any neuron can send a signal to any other by having lots of switching stations. There is plenty of time to do this, even with limited routing resources.

[chatham]

An article on an FPGA supercomputer at New Scientist

[Lawrence Phillia]

[Neil Hawkes]

Hi guys

Just a thought from a lay-observer.

I have been pondering the possible mechanisms of perception for a while in a DIY/hobbyist sort of way.

See http://positonicspage.blogspot.com/ for interest.

It seems to me that there is a difficulty in incorporating endemic feed-back (as opposed to feedforward) signals in any complex linear/analogue/continuous process - such as those ocurring in analogue neural networks - as the process rapidly becomes chaotic (in a bad way!) and that this is why there has to be a discontinuous element in neuronal signalling (the axon fires or not).

This is a matter of discontinuous output rather than encoding, obviously.