Slashdot Mirror
Kurzweil on the Future
dwrugh writes "With these new tools, [Kurzweil] says, by the 2020s we'll be adding computers to our brains and building machines as smart as ourselves.
This serene confidence is not shared by neuroscientists like Vilayanur S. Ramachandran, who discussed future brains with Dr. Kurzweil at the festival. It might be possible to create a thinking, empathetic machine, Dr. Ramachandran said, but it might prove too difficult to reverse-engineer the brain's circuitry because it evolved so haphazardly. 'My colleague Francis Crick used to say that God is a hacker, not an engineer,' Dr. Ramachandran said. 'You can do reverse engineering, but you can't do reverse hacking.'"
AI is our generation's flying car. It's what we see in the future, not what will be. Instead of the flying car, we got the internet. It isn't very picturesque (especially over at goatse.cx), but it is cool.
The future will be like that: something people aren't really predicting. Something neat, but not flashy.
Alternatively, the future will be the "inverse singularity" -- you know, instead of the Vinge godlike AI future singularity of knowledge, there could be a singular event that wipes out civilization. We certainly have the tools to do that today.
Use the Firehose to mod down Second Life stories!
Nonetheless there is something to what Kurzweil says, futurist (or in my language 'bullshit-artist') though he is.
The brain is probably impossible to 'reverse-engineer', not because of its evolution but because to come up with a brain you need to have 9 months in-utero development followed by years of environmental formation, nurturing and so forth, by which time the brain is so complex and fragile that analyzing it adequately becomes practically impossible.
I mean, take the electro-encephalogram (EEG). It gives us so little information it's laughable. Electrical signals from the cortext mask those of deeper structures and still we just end up with an average of countless signals. Every other form of brain monitoring is also fuzzy. Invasive monitoring of brain function is problematic because it damages the brain and the brain adapts (probably) in different ways each time. Sure, we can probably get some of the information we are after, but the entire brain is, I would suggest, too big a task.
But we can use the same principles that exist in the brain to mimic its functionality. But it ultimately is a new invention and not a replica of a brain, even if it does manage to demonstrate consciousness.
Do it yourself, because no one else will do it yourself. [beta blockade 10-17 Feb]
When I was in college in the '90s, our EE lab had just start experimenting with combining FPGAs with a genetic algorithm to model a non-linear function. Setup: pass in hundreds of random control streams of 0's and 1's that set up the logic of the FPGA, feed inputs and measure output pins, compare against a desired, then use the genetic algorithm to pick the "winners" that correctly modeled the function. The algorithm would randomly combine winners, then feed that back into the control stream, rinse and repeat until you have the "best" stream.
After that, the researchers took the control stream, mapped it back to find out which logic gates were activated / deactivated / multiplexed to route to one another. What they found was that there was no direct data path from the input to the output ! so how on earth were the output pins being generated?
What we were then told was that, somehow, the FPGA control pattern had created loops in certain parts of the circuit that was inducing current in the neighboring bus lines, like little radios and receivers. Totally non-intuitive, but mathematically it worked.
That is what I expect to see when we finally decode the human brain -- an immensely complex network of nodes whose linkages to one another were created in real-time using whatever resources were available to the "trainer" proteins at the time. No two individuals will encode the same event the exact same way: not the same locations in the brain, not the same method of linkages, or the number of linkages.
This is why I see the "singularity" not as a machine that we can walk into, have our brains scanned, and bam our consciousness can be copied to a computer. I think that every individual that "transcends" will have to do it incrementally, gradually replacing and extending the nodal functions that make up the brain. The brain needs to replace its neural network mesh with electronic blocks, and do it in such a way that the mesh's functionality is maintained while the material that makes up the mesh is replaced.
Over a period of time, there will be no more "meat mesh" and your conciousness would be transcended into a medium that we know can be copied, backed up and restored. And when that happens, well, our whole concept of what makes up a "person" would need to be redefined.
-- Scott
It might be less low hanging than most people think. Most predictions I've seen for, basically, "OMGWTFBBQ, computers are gonna be as intelligent as humans" are based on, basically, "OMGWTFBBQ, we'll soon have as many transistors on a chip as there are neurons in a human brain." Especially marketing depts love to hint that way now and then, but they're not the only culprits.
Unfortunately,
1. A neuron isn't a transistor. Even the inputs alone would need a lot more transistors to implement at our current technology level.
An average brain neuron takes its inputs from an _average_ of 7000 other neurons, with the max being somewhere around 10k, IIRC. The vast majority of synapses are one-way, so an input coming through input 6999 can't flow back through inputs 0 to 6998. So even just to implement that kind of insulation between inputs, you'd need an average of 7000 transistors per "silicon neuron" just for the inputs.
Let's say we build our silicon transistor to allow for 8k inputs, so we have only one modul repeated ad nauseam, instead of custom-designing different ones for each number of inputs between 5000 and 10000. Especially since, we'll see soon, that number of inputs doesn't even stay constant during the life of a neuron. It must accomodate a bit of variation. That's 2^13 transistors per neuron just for the inputs, or enough to push those optimistic predictions back by 13 whole Moore cycles. Even if you believe that they're still only 1.5 years each, that pushes back the predictions by almost 20 years. Just for the inputs.
2. Here's the fun part: neurons form new connections and give up old ones all the time. Your brain is essentially one giant FPGA, that gets rewired all the time.
Biological neurons do it by physically growing dendrites which connect to an axon terminal. A "silicon neuron" can't physically modify traces on the chip. You have to include the gates and busses that switch an input to another nearby source from thousands available outputs of another "neuron". _Somehow_. E.g., a crossbar kind of architecture. For each of those thousands of inputs.
Now granted, we'll probably figure out something smarter out, and save some transistor for that reconfiguration, but even that only goes so far.
There go a few more Moore cycles.
4. And that was before we even get to the neuron body. That thing must be able to do something with that many inputs, plus stuff like deciding by itself to rewire its inputs, or even (yep we have documented cases) one area of the brain decides to move to a whole other "module" of the brain or take over its function. It's like an ALU deciding to become a pipeline element instead in a CPU, because that element broke. In the FPGA analogy, each logic block there is complex enough to also decide by itself how it wants to rewire its inputs, and what it wants to be a part of.
There are some pretty complex proteins at work there.
So frankly even for the neuron body itself, imagining that one single transistor is enough to approximate it, is plain old dumb.
5. And that's before we even get to the waste we do with transistors nowadays. It's not like old transistor radios, where you thought twice how many you need, and what else you could use instead. Transistors on microchips are routinely used instead of resistors, capacitors, or whatever else someone needed there.
And then there are a bunch wasted because, frankly, noone ever designs a 100 million transistor chip by lovingly drawing and connecting each one by hand. We use libraries of whole blocks and software which calculates how to interconnect them.
So basically look at any chip you want, and it's not a case of 1 transistor = 1 neuron. It's more like a whole block of them would be equivalent to one neuron.
I.e., we're far from approaching a human brain in silicon. We're more like approaching the point where we could simulate the semi-autonomous ganglion of an insect's leg in silicon. Maybe.
6. And that's before we get to the probl
A polar bear is a cartesian bear after a coordinate transform.