Slashdot Mirror
New "Wet Computer" To Mimic Neurons In the Brain
A new type of "wet computer" that mimics the actions of neurons in the brain is slated to be built thanks to a €1.8M EU emerging technologies program. The goal of the project is to explore new computing environments rather than to build a computer that surpasses current performance of conventional computers. "The group's approach hinges on two critical ideas. First, individual 'cells' are surrounded by a wall made up of so-called lipids that spontaneously encapsulate the liquid innards of the cell. Recent work has shown that when two such lipid layers encounter each other as the cells come into contact, a protein can form a passage between them, allowing chemical signaling molecules to pass. Second, the cells' interiors will play host to what is known as a Belousov-Zhabotinsky or B-Z chemical reaction. Simply put, reactions of this type can be initiated by changing the concentration of the element bromine by a certain threshold amount."
Wow, talk about carbon bias!
Can you cite a reason why silicon-based systems shouldn't be as capable carbon-based ones? Silicon-based have developed at a blistering pace as compared to the carbon. (Though I admit that they have the advantage of actually having intelligent designers . . .) I mean, life has a head start of a few billions of years!
-Peter
Then you'll get computers with chemical imbalances. In other words suffering from depression. It gives the phrase "the computer is down" a whole new meaning.
As a preemptive strike:
http://en.wikipedia.org/wiki/Marvin_the_Paranoid_Android
putting the 'B' in LGBTQ+
I fail to see why you need chemical based computers in order to construct artificial intelligence.
One could build the system out of tinker toys and achieve the same results, at different speeds and different costs.
There is nothing that signifies intelligence which is provided by one construction method that is not present in another. Electrical, Optical, Mechanical, Chemical, Pneumatic... They are all just a means to an end.
Sig Battery depleted. Reverting to safe mode.
I can't speak for the GP, but I think we'll exploit properties we don't fully understand (say, by growing neurons on a grid that interfaces with them) much faster than we'll be able to translate those properties into other systems.
Is a lot like building an abacus using 3-D software and then manipulating your 3-D abacus to add 1 plus 3 to get four while chewing away millions of computational cycles...
We need a better way to simulate the effect of a neuron without having to re-create everything down to the last protein and lipid in a nerve cell....
Tsukasa: All I really want, is to be left alone...
Many people bring forward this idea. I think it stems from the fact that "traditional" AI (which has only really been around for 60 years or so) has not yet yielded a "sentient" computer. People feel that this somehow means traditional AI, and even our whole computational model can't yield sentience. They attribute intelligence to the fabric rather than the logic it implements. I think these people fail to realize that whatever computation biological brains implement, we could simulate it on traditional computers, *if we even knew what is being computed*. The problem is that, so far, beyond the first layers of our visual system, and some very simple systems, we know not much about the way the brain is connected. However, from what's been discovered in neuroscience, it seems pretty clear that the early layers of the visual cortex perform simple convolutional operation that do not involve quantum physics, or fancy shmancy things we couldn't do *more efficiently* with silicon.
The human brain is very complex, but given enough time, we very well might get to understand what makes us sentient and be able to replicate it in a computer. My personal opinion is that the brain is full of specialized hardware that has evolved over a very long time, and helps us to specific tasks (eg: facial recognition, hand-eye coordination, obstacle avoidance, language decoding), with a very powerful abstraction logic built on top (the stuff that "makes us sentient"). This abstraction logic is possibly very complex, and perhaps too difficult for us to conceive of at this time. If we are to learn anything from the rest of the brain, most of this logic probably focuses on transforming perceptual information into a form that makes it easy to reason with. On top of this, we probably again have specialized mechanisms, to do things like deduce causal relationships and generate hypotheses or semi-random associations of concepts (creativity).
The reason the "traditional AI" camp hasn't succeeded at making sentient machines are multiple, but I would sum them up as follows:
1) They have mostly given up. You probably can't get funding for claiming you'll come up with HAL9000, you'll sound like a wacko. Current AI research focuses simple learning problems (i.e.: supervised learning, reinforcement learning).
2) The approaches tried in the past focused purely on formal logic, which, as we now know, works badly in open-ended environments. For it to work well, the properties of the environment have to be simple, restricted and well-defined.
3) Supervised learning, unsupervised learning, etc., will not yield sentience. These approaches, which may actually exist in the brain, are good at solving problems of limited scope only. Our brains are not big wads of neurons performing a single computation. They are much more intricate and integrate many specialized components.
The "right" approach to AI is probably an overall approach, integrating many existing techniques into one system. Perhaps an "engineering" approach to AI would work better. Focus on constructing it and then refining it, as opposed to developing an overall theory of how it will work first and trying to reduce it to its simplest component. We already have computer systems that do speech synthesis, speech recognition, facial recognition, depth perception, 3D model reconstruction, etc. We also have unsupervised learning, supervised learning, reinforcement learning, fuzzy logic, knowledge bases, automatic theorem provers, etc. It should be possible to build a non-completely stupid AI, if one combined all these techniques in the appropriate way. How to connect them, however, is probably where the true AI problem resides.