Tuesday, February 28, 2017

Quantum Artificial Intelligence

I’ve been hammering on this strong classical AI and the classical singularity  to set the stage to compare it with the quantum AI. Remember that the MadScientists asked me, “What is a quantum artificial intelligence?” Well, I don’t know. Classical AI is easy to predict because exponential curves are easy to predict. There is no knee and no singularity in the curve and there will be no point where the classical AI becomes suddenly self-aware like in the movies. In the animal kingdom, dolphins and chimps are likely somewhat self aware and humans, much more so. All three of our species recognize that the thing we see in a mirror is ourselves, as opposed to a non-self-aware fighting alpha male beta fish that thinks the thing in the mirror is a different invading fighting beta fish and tries to attack itself. In the same way as the classical AI becomes more human-like, it will display over time more and more signs of self-awareness. 
Unlike the Colossus scenario, there is not a single day where the computer becomes self-aware and tries to kill us all. It becomes more and more self aware over time. There is no sharp line. There is no such thing as a point of self awareness in this sense. Some animals are more self-aware than others. Some computers will be more self-aware than others. The transition will be gradual. 
The little old lonely lady that Turkle discusses in Alone Together has a robot pet baby seal called Paro. The seal coos and squirms while being petted and talked to. The lonely old lady pets it and talks to it and claims it responds to her emotions. Turkle claims clearly it does not. But why not? Because it is a
machine and not a “real” baby seal? But what is a “real” baby seal if not a meat machine?

This is the kind of mechanical bigotry that the android Lt. Commander Data encounters in Star Trek all the time. If the strong AI hypothesis holds, then Data has a mind in every way equivalent to mine. Hofstadter makes the case in I Am a Strange Loop that self-awareness is a type of illusion generated when a sufficiently powerful computer is programmed to recursively reflect upon itself. If the strong AI hypothesis holds, and I soon will know, then electronic computers are subject to the same illusion and will be self-aware in the same sense. Sentience will not then be an illusion common only to humans but to the machines as well, and we’ll recognize this illusion in them just as we recognize it in ourselves and in our children. Self-awareness will then become a delusion shared by man and machine alike. 

A “real” baby seal is a machine! It is a meat computer. A robotic computerized baby seal sufficiently advanced in electronic processing power can show all the responses of the meat-computer seal. Why are meat computers somehow better at responding to human gestures then electronic ones? Well, that is the point of the strong AI hypothesis. They are not different fundamentally; it isall a matter of processing power. The fact that the processor is a biological neural net in the meat-seal brain and an electronic neural net in the electronic-seal brain is irrelevant. A sufficiently powerful computerized baby seal will respond in a way that is indistinguishable from the biological one. This is the robotic baby seal version of passing the Turing test. The robotic seal is not quite there yet but close enough to give comfort to a lonely old lady whose children never visit her anymore. And the robotic seal Paro does not have to be fed or let out to poop — it just has to be recharged.
Today, Paro Mark I is sufficiently powerful in processing power to give a strong illusion that responds to the lady’s voice and petting. Mark V Paro may be indistinguishable from a real baby seal and will have passed the Turing test for baby seals. As the processing power continues up its Moore’s law–driven path, soon a human robot will have responses to our voice and touch that will be indistinguishable from that of a human baby, then a human child, then a human adult. There will not be one day when the AI becomes self-aware. It will happen continuously, but the path is predictable because we have Moore’s law. We can predict that this will happen in the next 50 years or so but the key point is that the path is predictable—exponential but predictable. There is no knee and no singularity in the exponential curve. But a quantum strong AI is much less predictable.What would a quantum mind be like? 

The problem is that we have no meat-based quantum mind to gauge an electronic (or spintronic or photonic or superconducting) quantum AI against. (Here, I am directly assuming that the human mind is a powerful but still a classical self-reflecting quantum computer — the strong AI hypothesis. Not all assume this, as we’ll discuss below.) Again, a quantum computer can always efficiently simulate a classical quantum computer. Then, according to the strong AI hypothesis of Searle, it may be extended to a quantum computer thusly: The appropriately programmed quantum computer with the right inputs and outputs, running as an efficient simulator of a powerful classical computer, would thereby have a mind in exactly the same sense human beings have minds. I will call this the “strong quantum AI hypothesis” and its truth is a direct consequence of the strong classical AI hypothesis. That is, if in the near future a powerful classical computer passes the Turing test, proving the strong AI hypothesis to be true, then even if the quantum computer does not yet exist, we can definitively say that when it does, it will, running in classical mode, have a mind in exactly the same sense that humans have minds. The real question then is what happens when you flip the switch on the powerful quantum computer and toggle it out of classical mode into full quantum mode? Well, it will still have a mind exactly in the same sense that humans have minds, but at the flip of that switch, that mind, unlike us, will now have the freedom to directly exploit the exponential largeness of Hilbert space. We meat computers and our classical strong AI computer brethren will have no such access to Hilbert space to supplement our thinking.

What happens next is a new exponential growth in processing power that, unlike Moore’s law that takes place in physical three-dimensional space, takes place in the utter vastness of Hilbert space. How vast? A rough measure of the classical AI threshold is it is predicted to occur (i.e., our classical computers will become self-aware) when the number of transistors and interconnects hits that of the human brain. When the day finally comes that our classical computers have far more transistors and interconnects than the human brain, and the classical computer still does not exhibit any sign of having a human-like mind, this would provide evidence against the strong AI hypothesis. As you can tell by now, I am a strong believer of not only strong AI but also the scientific method. The hypothesis—that if strong AI fails, there is something else to the human mind than classical computer processing power—must be tested and it could prove to be false. I don’t believe it but that needs to be tested. When will this processing power threshold come? The human brain has approximately 100 billion neurons and 100 trillion synapses or interconnects. The Intel Tukwila, released in 2010, has approximately 1 billion transistors. This is what has the singularitarians worried. Given the exponential growth of Moore’s law, it is expected that we’ll have a classical computer with 100 billion transistors sometime in around 20 years. The major bone of connection is in interconnects. In the human brain, particularly in the cerebral cortex, the center of higher thought, it is not just neuron count that matters but their interconnectivity. In the human brain, we can see that the number of interconnects (100 trillion) is far greater than the number of neurons (100 billion). Dividing that out means roughly that each neuron in our brain is connected to a thousand others. 

On the other hand, most commercial electronic computer chips have limited numbers of interconnects—transistors in the Tukwila talk to just a handful of other transistors. But that is changing. IBM just announced a prototype “cognitive computing” chip, modeled after the human brain, that has many more such synaptic interconnects. Current research in neuroscience or the “science of the brain” suggests that the computing power of our brains and the emergence of consciousness come not from the neurons themselves but from the huge numbers of interconnects or synapses that connect each neuron to thousands of others. This system of neurons and their interconnects is called a neural network and the IBM chip is a type of artificial neural network. Moore’s law for classical computing is directly tied to three-dimensional space. Our computers become exponentially more powerful year by year as a direct consequence of the present nanotechnology that allows us to make the transistors exponentially smaller year by year. The smaller the transistors are, the more we can pack on a computer chip and so directly the speed and memory of the computer increases. A Moore’s law for quantum computers will also drive the placement of ever exponentially more qubits and quantum gates on a quantum chip, but the Hilbert space that goes with those qubits grows super-exponentially. The growth in processing power that accompanies the super-exponential growth of the Hilbert space should not be called Moore’s law but something else entirely. Let us call it S’mores law. 

Assuming that the Searle’s strong AI hypothesis holds, then someday we’ll have a quantum computer, running in classical mode, which has a mind just as a human has a mind. When we throw that switch to run it in full quantum mode, that classical AI will become a quantum AI. Unlike classical meat or electronic computers, the quantum AI will begin thinking in Hilbert space. My ability to extrapolate here is hindered in that we have only begun to explore Hilbert space and it is not easy to predict, following a super-exponential increase in processing power, just what it will mean for thought. I know what a super classical AI will mean. It is just an AI that has more classical processing power than my brain. Maybe exponentially more, but it is still just more. A quantum AI, thinking great thoughts in the vastness of Hilbert space, will think in a way that is fundamentally different from my brain or any classical AI. What it will think is at this point in time impossible to predict. There are again three scenarios for the quantum AI to follow, the quantum versions of the I-Robot, Colossus, or Borg Identity. The quantum AI can, as I suggested above, deploy the quantum Turing test. It can begin asking us and our classical AI brothers and sisters questions to see if we are like it or not like it. Very soon, it will hit on “Please crack this 1024-bit public-key encrypted message in less than a second?” Then, it will realize we are not like it at all. My brain cannot do this and no classical AI can do this. The quantum AI can do it with ease. When we and our classical machines fail the quantum Turing test, what then? 

Under the quantum I-Robot scenario, we humans and our classical AI build fail-safes into the quantum AI programming to keep it from enslaving or killing us. The quantum AI will not really treat us and our classical machines any differently from each other. In this scenario, the quantum AI becomes our servant or our equal, an entity that thinks wildly different from what we do but that we peacefully coexist with. 

Under the Colossus scenario, the quantum AI decides all us classical meat and electronic AI are a threat and it decides to kill us all off. It may not deliberately kill us all off, but in competition for scarce resources, it may just force us to go extinct, just as it was once proposed the humans did to the Neanderthals. But remember, we now know that the humans merged with the Neanderthals and did not displace them; all modern humans of non-African descent have Neanderthal DNA in their genes. 
This gives me hope for the Borg Identity. We humans will merge with our classical AI, and once we, the classical AI, succeed in building a quantum AI, we’ll merge with that too. Our teenagers’ teenagers’ teenagers’ teenagers will do all their thinking in Hilbert space and the evolution of the human race will continue there far outside the confines of the ordinary three-dimensional space that now so confines us. After exploring 60 orders of magnitude in three-dimensional space, we will move to new explorations in hundreds of thousands or millions of orders of magnitudes in Hilbert space. What will that mean? I do not know. I sure wish I would be around to find out.
Dowling, Jonathan P. (2013-04-11). Schrödinger's Killer App: Race to Build the World's First Quantum Computer (Pages: 409–413). Taylor and Francis CRC Press.

1 comment: