This is a publication of the Discovery Institute, a politically conservative think tank which advocates intelligent design and has a department called the Center on Human Exceptionalism.
Note the sneaky non sequitur at the end, which of course this being the Discovery Institute was the whole point of the article:
>Maybe it’s worse than they fear. Perhaps the brain doesn’t and can’t operate like a piece of machinery at all. And vice-versa.
Just because the brain may not work like an 'ideal' neural network running in a computer simulation, that doesn't in any way say that it couldn't be duplicated using a non-'ideal' neural network simulation.
It may just mean we need to adjust the form of the simulation.
Yeah, the discovery institute is just another creation science cover front. No science, just 'truth'. They try hard to look legit. And a fascinating fact is they are in Seattle of all places. It's a 'the earth is 9000 years old somehow' place and "let's cast suspicion on all science that contradicts the bible".
You’re implicitly suggesting that the brain “operates like a piece of machinery” and that we just need to adjust the form of “simulation”. The problem with this is that you’re also implying that the brain (and human mind) is a deterministic system.
The implication of determinism is that everything is predictable. Given a set of inputs, in a deterministic system whose rules we fully know, the output is entirely predictable. We could extend this determinism to the whole universe, and reduce the universe into a pure function 'f' with a some specific initial state 'x'. Starting with f(x), for every unit quantization of time, f is applied on its previous result --- f(f(f....f(x)....))). And that's reality. (If you believe in a Big Crunch / Big Bang cycle, the output of 'f' at some point cycles back to the original 'x'.) The initial state 'x', and the definition of the universe function 'f' _is_ existence -- the totality of all that is.
I don't have any proof that we are not living in a deterministic universe, but all I can say is that, philosophically, I find the idea of determinism to be extremely upsetting. It empties life of any meaning. Why even continue living if everything you do was pre-programmed at the beginning of the universe? Determinism implies that the values of 'f' and 'x' could be blamed for the Holocaust. Moral responsibility becomes meaningless, because everything is pre-programmed. I think a lot of people, if they deep-down believed in this sort of mechanical determinism, they would sink into a very deep deep depression. I struggled with this myself, as a teenager—and I’ve emphatically rejected human mind determinism.
Believing in free will on the other hand, is very empowering. Believing in the concept of free will can in and of itself imbue one's life with a certain degree of meaning, purpose and drive. It means what you choose to do, matters. Life is meaningful; we’re not just empty automatons executing our pre-programmed futures under illusion of choice / free will. We are sentient beings with free being.
From the point of view of physics, I think free will in biological organisms will eventually be proven to some degree. The fact that in the last century, practically every physicist has come to accept that there's a lot of unpredictable behavior in matter, that cannot be deterministically predicted, and can only be described with probability distributions, has been very reassuring. My theory is that free will emanates from somewhere in these probability distributions, and is amplified in living organisms.
This is interesting. You are basically saying that free will is the same as non-determinism. I don't think that's the case. For me, non-determinism is just randomness (quantum for example), and doesn't imply free will. Can you just say 'your actions are random, therefore you have free will '?
First, 'free will' in the physics sense doesn't really exist. Everything is particles, and particles don't have 'will'. Your body is constituted of particles that follow their own path and interact. This can be deterministic or not, it doesn't matter.
Then, you can believe that there is something that we cannot see that can be called 'free will', like a soul, but this is a religious belief.
But we can look at free will in the philosophical sense. People have a conscience, they are conscious of themselves, and of their own free will. Also, your actions are only a product of your own inputs (your senses) and your outputs.
I believe those are better arguments for defining free will. As free will doesn't exist in physics, we have to give it a definition. And I think this is true whether the universe is deterministic or non deterministic.
> No, I'm saying that determinism precludes free will.
> Free will is more than mere non-determinism / randomness.
Yes, I even think that free will is not really related to the determinism / non-determinism of the universe, as I said earlier. Free will doesn't make sense in physics.
That's why I don't agree that determinism precludes free will. You can define free will in an other way, that is compatible with determinism. That's my point of view, you can read more about it : https://en.wikipedia.org/wiki/Compatibilism. Your view is called 'Incompatibilism', ie determinism and free will are mutually incompatible.
At the end, I think that most of it is a question of definition, but we can all agree that free will in a deterministic world is not the same as free will in a world where people have souls.
It’s impossible to have actual non-determinism in a deterministic system. Non-determinism is emulated in software using random number generators. If you know the seed (the ‘x’) and the definition of the random function (the ‘f’), you can predict all outputs, analogous to my f(..(x)..) universe above.
There’s no programmatic random number generator that is truly random. (I mean random in the sense that it’s entirely unpredictable; not in the sense that the distribution of the random number generator covers all possibilities uniformly over many calls.)
Hardware random number generators are truly random (like the RDRAND instruction in x86 CPUs which is seeded by an on-chip entropy source). But if you’re relying on physical entropy, it’s not a deterministic system anymore. You simply can’t emulate non-determinism in a deterministic system.
You compute every possibility in parallel, and defer the selection of the 'true' result to some time-at-infinity. Once you've done this, realize that there's no difference within this system between selecting the 'true' result at random or using an algorithm, because the system would consider this result selection to be equally arbitrary in either case. (I.e., there's no way to prove that you've selected a specific result non-deterministically.)
A deterministic system emulates a non-deterministic one using (computational or data-) parallelism, and a non-deterministic system emulates a deterministic one using (temporal or spacial) aggregation.
We typically use random numbers to model non-determinism because it is computationally infeasable to track every possibility, but that is not the only way to do it, and nature doesn't appear to have this problem.
> there's no difference within this system between selecting the 'true' result at random or using an algorithm, because the system would consider this result selection to be equally arbitrary
This makes zero sense. Choices matter. Selecting one path could potentially result in the suffering of billions of people, and another path could result in the joy of the same billions.
> We typically use random numbers to model non-determinism because it is computationally infeasable to track every possibility
First, tracking every possibility has absolutely nothing to do with random selection. I don't know how you're mixing these two. We use pseudorandom number generators precisely because it is impossible to have true randomness within a deterministic system.
> You compute every possibility in parallel
I think you're alluding to some sort of multiverse theory -- and implying that since every possibility exists in parallel, choices are meaningless and nothing matters.
Except, we don't live in a multiverse -- a gobbledygook theory if ever there was one. Choices matter.
>We use pseudorandom number generators precisely because it is impossible to have true randomness within a deterministic system.
"Nondeterminism" and "true randomness" can be modeled in deterministic systems. Your ability to deterministically construct those notions in your own brain is a weak proof of that. If I say "elephant", and you imagine an elephant, then you've modeled an elephant. You have not created an elephant, I'm not claiming you will, and I'm not claiming that you will get true randomness from a deterministic machine. I'm saying you can model it, and thus reason about it.
>I think you're alluding to some sort of multiverse theory.
I am not. I am alluding to parallelism being a computational feature which usefully serves as a model of nondeterministic processes.
I'm not talking about choices, meaningless or otherwise, I'm talking about modeling the concept that is invoked by your use of the word "nondeterminism." (Though you can also model nondeterministic choices as parallel thunks.)
My point is that for every nondeterministic function, there exists a deterministic function which will produce exactly the same outputs. Thus one cannot argue that nondeterminism offers any additional capabilities (beyond computational efficiency.) This is effectively the same as saying that every parallel computation can be executed sequentially, or that every strict computation can be computed lazily.
Stochastic determinism is still determinism, in the philosophical sense. On the other hand, certainly, the probabilistic (and metabolically/computationally bounded) nature of the mind is an important part of why it doesn't seem to fall into certain "paradox traps" the way a nonstochastic axiom schema might.
You should probably go right up on compatibilism about free will, because at this point, you're re-litigating a debate that hasn't been considered philosophically interesting or nontrivial for ages.
Just build a non deterministic AI. Free will and true AI are not (necessarily) mutually exclusive.
"Free will" is a vague concept anyway. What does it really mean to make a decision? It already is heavily weighted by your knowledge and experience, everything else would be "random".
A smart (smarter than us) machine shouldn't have "free will" anyway, as all choices should be exhaustively evaluated for their pros and cons.
> he problem with this is that you’re also implying that the brain (and human mind) is a deterministic system.
This is not a problem. This is a conclusion from the assumptions. Just because you don't like the conclusion does not mean it's incorrect.
Does a transistor have free will? does a protein? a bacteria? A sponge? A mouse? A dolphin? A human?
Another point of view of "free will" is that it's what we call a system that is so complex, has such good feedback loops and is so nuanced that it is capable of influencing itself. That is not exclusive with determinism.
This argument has been going on for a while Einstine said:
"der alte herr doesn't throw dice" or "The old man (God) doesn't throw dice"
His comment was in reference to Quantum Mechanics which is based on stochastic functions and make this deterministic approach invalid. This was one of the instances where Einstein was wrong.
Something working with neural networks has helped me to come to terms with is that we are probabilistic creatures, and even things we call facts are things we simply feel very confident about. Science itself is a process of increasing confidence in an outcome, not an exercise in certainty.
The fact that an artificial neural network (which very roughly mimics human brain) is probabilistic in nature does not have any implication for Science as a whole.
Science can make predictions that are extremely accurate (think ten significant digits) and the uncertainty on that prediction (i.e. the eleventh significant digit) does not have anything to do with neural networks or us being "probabilistic creatures", whatever that means.
Even in your ideal experimental scenario, the laws of probability state that before the 10,000,000,000th such study is completed, we will see a case where the findings turn out to be incorrect.
We learn by experience, and we build weightings, and both our gut reactions and careful conclusions are simply the ones which win the confidence ratings. Logic may be an exception by virtue of it's mathematical nature, but any facts involved are not, and we fall back on judgement calls most of the time even when we think we're being logical.
> Even in your ideal experimental scenario, the laws of probability state that before the 10,000,000,000th such study is completed, we will see a case where the findings turn out to be incorrect.
And yet by means of statistics one can model the errors, and obtain an experimental measurement with 10 significant digits.
And, guess what, it agrees with the theory! Ten. Significant. Digits.
> Logic may be an exception by virtue of it's mathematical nature, but any facts involved are not, and we fall back on judgement calls most of the time even when we think we're being logical.
This article is rather confused. First, just because processing is distributed doesn't mean that there aren't different areas of the brain that are more involved in certain things. The visual cortex is pretty clearly for vision for example.
Second, they seem to be saying that AI is not like the brain because it is a "switch" (what?) and doesn't have distributed processing/representations like the brain? Or something? It's hard to know what they're trying to say because it doesn't make any sense.
> Thus they determined that the brain does not operate like a set of switches, as we used to think
Nobody thinks that. Not even in neuroscience. Karl Lashley's experiments with cats happened well before the 50s and he concluded that there is no specific place where memories and learning are formed. We know that some areas specialize in mainly doing one thing or another
Yes. So many scientists, so many biases, much still to discover/understand. I wonder if we will be able to find some solid info on what's going on. I like Jeff Hawkins discussion surrounding this idea, or David Chalmers idea of the hard problem. We have tons of info/facts about what's happening in the brain physiologically, locations of things, neutral pathways, etc. We don't have a grasp on consciousness though. But, Chalmers, Hawkins, Dennet, ...... All have strong beliefs. It's a curious fundamental mystery, and so wonderful.
Not a mystery at all, unless you start with the premise of the "Human Exceptionalism" and as jolux rightfully points here, the publisher Discovery Institute has even a whole department called "the Center on Human Exceptionalism."
That's how they start from the false premise and come to the false confirmation of that false premise with a religious answer that it's "a mystery."
Interestingly, the editor of this particular publication writes in his article that he is aware that one doesn't have to start from such a false premise, but that they go "way beyond this":
"Consciousness, for example, is defined in the dictionary as being aware of your surroundings. A robot with environmental sensors is conscious by this definition. In our discussions, the topic of AI consciousness is way beyond this."
At least he doesn't use in that text the direct "it's a mystery" approach. But looking at the titles of the articles there where he's an editor we see:
"Possible Minds?: But What If the Minds Are Impossible?"
"New Evangelical Statement on AI is Balanced and Well-Informed "
To those who haven't yet read the article, don't bother. Here's a 15-second rendition of the argument to save you the time:
A rat study showed most of a rat brain was used to process information from touch. This does not match ‘function localization’ models of the brain, and might be relevant to brain injury outcomes. There is also debate to what extent, if any, neuron growth happens in adulthood. Therefore we should be suspicious about neuromorphic computation and maybe the brain isn't substrate independent.
No, I'm not skipping a step with ‘therefore’, it really is precisely that abrupt, arbitrary, and poorly-reasoned.
It's not an in-depth article, but it does make an important point: That we actually understand very little of how our brain really works.
People here on HN used to have a lot of discussions about uploading brains, neuro-preservation, simulating brains, etc. The implicit assumption was always that we already have a pretty solid understanding of how the brain works (it's just a neural network) and we just need to solve a few minor problems (just take a snapshot of all the synapses and we're good to go!).
The truth is, of course, that we have no clue if a neural network even is a good model beyond the most simple circuits (like certain reflexes), and scientists are still scratching the surface.
I mostly agree, but with two caveats. We do actually know that ‘neural networks’ are poor models for the brain. Backpropagation is unphysical and brains aren't acyclic graphs, for example. What there's uncertainty about is the degree by which they share principles. A glider and an eagle are pretty darn different, and a glider is a terrible model of an eagle, but they do at least utilize the same laws of fluid dynamics to fly. So the question holds for neural networks.
As to neuropreservation, it's important to note that we don't need to understand the brain in order for it to be a good idea. What we care about is whether there will be a future wherein someone does, and whether the technique preserves enough mental structure to allow the brain to be recovered by that somebody.
Yeah, and it fails to acknowledge that there is a long and nuanced body of work in neuroscience concerning the degree of localization vs distribution of function going back at least 100 years (see Karl Lashley, Donald Hebb)
Neuro-morphic, having the form of neurons. Recently the buzz refers to circuits that perform neural network operations (the AI kind) using hardware properties directly, storing weights and performing inference or backprop with the computing substrate, which is potentially more efficient or faster than using SRAM and multipliers. It can also refer to hardware directly attempting to model biological neurons, which is more far-field research than the former.
Substrate independence denotes a property being separable from its physical implementation. One could say Alan Turing showed that computation is substrate independent, since many machines all do the same sort of computation. It's often debated whether consciousness is substrate independent, whether a piece of silicon just simulating a human brain will actually have an internal experience of its own thoughts. People who believe in souls, dualism, or panpsychism tend to think it isn't (to different degrees), people who believe in reductionism or emergentism tend to think it is. The authors of this piece go so far as to say the stuff the brain does might not even be reproducible outside of flesh-and-blood, regardless of whether such a reproduction would have internal experience.
If I recall my terms correctly, substrate independence is the idea that creating intelligence is not dependent on using neurons, but that if you could model the required functionality of neurons in another substrate, an intelligence could also be housed in that substrate.
> Neuromorphic computing
Is basically modelling neurological architectures in computers.
