Brain vs. Deep Learning (2015)
timdettmers.com
timdettmers.com
I think in a sense this is well-established when looking at real neurons, but several times in the article he uses phrases like "shown to be important for information processing", and that's where I get off the boat a bit. When you're saying that it's so important for information processing that it warrants a 1000x or more increase in the computational power necessary to implement an algorithm, I think it's necessary to dig into what the actual work being done there is, not just that there's some non-trivial transformation. A lot of interesting and extremely tough to model fluid and chemical dynamics are in play when I drink too much water and have to pee, but that doesn't mean that we need to understand them to build a waste disposal system using pipes.
In particular, does the within-neuron processing actively tune itself based on the data it processes to an extent on-par with inter-neuron connections (in which case the argument that it's fundamental to the learning process would hold a lot more weight), or is it mostly static? I think a lot of us consider "important for information processing" to mean "is a meaningfully dynamic parameter involved in a learning algorithm", rather than an accidental shmearing of data.
I'd really love more info on what the actual processing that's happening is.
Does it tune itself or is static? It definitely tunes, but it can be meaningful to the algorithm or ignored.
I'll just leave this here: https://en.wikipedia.org/wiki/Gene_regulatory_network
Inside any cell there is a system that works like a neural network - the gene regulatory network. Each gene acts like a neuron, with chemical inputs and outputs. That would make the processing power of any cell on par with that of a small neural net.
I'd be willing to bet that once we understand the basic functioning, we'll be able to build a working brain using a few different types of neuron which operate on relatively simple rules-of-thumb, arranged in a few different basic structures which are internally relatively homogeneous.
We absolutely need to start doing what you're suggesting, and figuring out how to derive useful basic structures, represent them, and construct networks using those as the building blocks. My prediction is that's where the field will go over the next 5-10 years, a much deeper dive into how to specify connectivity and usefully control it than has been done so far (we're still using fully connected layers, for the most part, which we know is not a scalable approach as we go from thousands to millions of nodes).
To pick one point of disagreement,
“We do not need as much computational power as the brain has, because our algorithms are (will be) better than that of the brain.”
I hope you can see after the descriptions in this blog post that this statement is rather arrogant.
Machines are already better than brains at many cognitive tasks of practical interest. Believing that we'll continue to find "tricks" to allow computers to outperform brains on useful cognitive tasks, despite the brains' much greater complexity, seems like a perfectly sober and conservative prediction.
If I had to advance my own pet reasons for discounting the likelihood of a technological singularity, here are my top two:
1) It's a more challenging case of the general Fermi paradox. Show me the Hubble images of the computronium Dyson swarms. If it takes less than a century to go from the first transistorized computers to superintelligence, and superintelligence is as prone to run amok as Bostrom/Yudkowsky think, signs should already be visible from Earth.
2) You need experiments to validate scientific models. Even if a machine-intelligence could think a billion times faster than a biological intelligence, it couldn't complete experiments a billion times faster. Technologies that act on the material world will improve sublinearly with respect to thinking/computing power, for at least this reason and probably others as well.
AlphaGo and even labeling are indeed milestones in superhuman performance, and I believe Terence Tao's view that AI is a moving target is relevant here.
---
1) Regarding the Fermi paradox, I haven't read too much technical or scientific work on it (it's highly speculative anyway), but I find Isaac Arthur's videos lay the basic ideas pretty well: https://www.youtube.com/watch?v=oIva_60l3ww&t=1016s
It becomes pretty convincing that technological intelligent life is an incredible coincidence. There might not be any huge "filters", as they are called (stages which reduce the probability of observing TIL), such as a superintelligence extinction event, but an enormous sequence of minor filters, ranging from low probability of a planet with adequate conditions, to low probability of actual technological development, to low probability of occurrence in our light cone (we can only see fairly young distant galaxies), and limitations to the visibility and spreading velocity of such civilizations.
There is a wide range of parameters such that this does not contradict the generalized Mediocrity principle, such that there is probably more than one TIL in the universe, but they're few and far from each other in space and time.
2) I don't follow. This objection is only valid for discovering new laws of physics -- technological development can happen even with fixed knowledge of basic physical laws. At this stage it's not clear that even continued refinement of physical models. For example, the discovery of quarks certainly helped understanding nature, but it hasn't led to any direct technological applications due to quark confinement, and the fact that the particles are just too small. Neutrinos are another conceptually important discovery that doesn't really have applications due to low interactivity with matter.
No, experimentation is also important for translating fundamental scientific knowledge into working technologies. Consider the problem of building a fusion reactor that generates net electricity. You don't need new laws of physics, but you can't make progress with simulations alone. You have to experiment to find out if devices behave in reality the way they do in your model.
I think a greater issue is the general misconception of intelligence as something of a magical attribute, and even more so "superintelligence", and even more so "recursive-self-improving superintelligence" (RSIS for short?). There are not only various limits to hardware, some of which discussed in the article (size of atoms, capacity of communication between parts), but also there are limits to software. For example, while it seems that an RSIS could really solve any problem it wanted, Turing's almost century-old Halting Problem solution already proved that no such computer exists. There's no algorithm that takes an arbitrary conjecture, e.g. the Riemann hypothesis, and outputs either {yes,no,malformed_problem}. In fact the general expectation of what AGIs can do (think creatively, solve arbitrary problems, replace mathematicians, replace programmers) are things that we ourselves can't really do. We solve problems using a small set of heuristics and lots of trial and error, with no success guarantee.
Another fact is that there are many asymptotically optimal algorithms that are already know, and the trivial fact that almost all tasks performed by this super-intelligence (many of which are ordinary tasks we already do, analogous to maybe sorting, database queries, and search engines) will just improve constants on near-optimal algorithms.
A god-like creature those AIs will be not.
Current day philosophy is like current day ML - detects correlations but can't run experiments to filter out the bad ones, so they (philosophers) are stuck with a combinatorial explosion of theories. They key component missing is a good enough simulator, both at physical and abstract level.
Yann LeCun saying the main problem in AI is how to make a predictive model of the world (a simulator): https://youtu.be/cWzi38-vDbE?t=1933
> I hope you can see after the descriptions in this blog post that this statement is rather arrogant.
The brain does more than computers: the brain builds its internal structure all by itself. When was a computer able to evolve from a bunch of transistors laying on the table? Also, it uses very little energy and is resilient for 80-100 years, compared to the computers of today, its apples and oranges.
> Show me the Hubble images of the computronium Dyson swarms. If it takes less than a century to go from the first transistorized computers to superintelligence, and superintelligence is as prone to run amok as Bostrom/Yudkowsky think, signs should already be visible from Earth.
What if the AGI will prefer to build virtual worlds and societies of virtual agents instead of grand space domination ? Even humans prefer games to reality nowadays (or a large percentage of us do). If we're part of such a sim, then it would explain the lack of external signals from extraterrestrial aliens or AGIs.
The probability that AGI will appear and create amazing simulations is much larger than that of picking up signs of life in the vastness of space. Also, by replacing physical with sim we can do all sorts of things such as use less energy, recover from any accident, hack our own brains/minds, even immortality.
This is against the 13W that the brain uses: https://www.scientificamerican.com/article/thinking-hard-cal...
So the brain has a million times more the processing power and uses a million times less power.
A commodity smartphone is millions of times more powerful than all of NASA's combined computing in 1970 and more energy efficient.
A factor of 10^6 is maybe achievable with a specialized quantum computer in the mid-term future ?
>when was a computer able to evolve from a bunch of transistors laying on the table
These two statements are not really connected. Brain builds its internal structure in a sense, it changes with learning for example. But brain did not evolve by itself. It's all thanks to the force of evolution.
So it boils down to the question: "Can we do better than evolution?", so we need to understand what evolution is. It's just random mutations that make organism better than it was before. If these random mutations make it worse, organism gets lower chances of survival and usually just dies. Essentially, evolution is a method of design by brute force. And brute force works, if you have a lot of time, but it's common sense that intelligently designed algorithm can do a lot better. So we can answer the question now with "Yes, most likely we can do better."
tl;dr doing pure Deep Learning (also known as connectionist) models of the brain limits your tools in a bad way. Using tools from Dynamicism and Symbolicism is better. As proof, check out Spaun, the world's largest functioning brain model.
Note: I mostly just disagree with him philosophically, in terms of his reasoning because it's overlooking some evidence. Don't really have an option on his conclusion. Probably agree with him more than I disagree with him.
But yeah 2080 is within my daughter's lifetime. My 1950s house I bought will probably still be standing. I think it also coincides with Elon Musk's guess for when there'd be 1M people living on Mars.
He wrote this in 2015. Was he correct? I don't really know where to start researching something like this; is there anyone here familiar with the field who could comment on it?
Note the "quantum" here is not a good thing (like the theoretical "quantum speedup" possible for certain computations on quantum computers), but a bad thing: imagine you want to send current down this wire, but the current is jumpy and often leaks out of the wire to a neighbouring wire thus causing errors in computations.
We are so used to this framing, that it may seem foreign. Others will argue, that in principle it is a problem of computability but in extreme almost every problem is, but that is not necessarily how we frame other problems. I'd say "intelligence" is a control problem (as in controlling a robot). This framing, though subtle, makes the entire problem quite different. You no longer talk about computability but you talk about survival in "high temperature thermal bath" (or otherwise called "physical reality"), full of unpredictability and dangerous stuff.
When you frame it like this, it is clear we have not even began to address the problem properly, not to even mention solving it.
Perhaps treating AI as a derivative of the control branch of computation could practically help speed up progress in some areas, but it shouldn't fundamentally change its nature.
We get all the great marvels of deep learning, but robots remain stupid as bricks over 30 years of Moore's law. To me this (Moravec's paradox) is a signal that we are doing something wrong, and typically we do things wrong when they are not framed properly.
Some of what you mention are what I consider non-well-defined philosophical problems, which really have no bearing on the ability to create algorithms that could pass the Turing test and are outside of the scope of a discussion on the practicality of creating AGI.
I see from another post of yours that you are religious, which is a common personal reason to assume a form of dualism as at least central tenets in the Abrahamic religions seem to require it. Do you see dualism necessary for your position of human intelligence being incomputable, and if so could you reference an argument you would find compelling also for readers whose religion (or lack thereof) does not necessitate dualism?
If you don't see dualism as necessary for the possibility of human intelligence being incomputable, could you open up your thinking on this?
I apologize for being so forward, but I worry some might read your comments without realizing that they might rely on unshared assumptions.
"Is it true there is no truth?" is not even close to a compelling argument, there are rigorously defined notions of truth within any of the eminently computable logical frameworks that we use.
If you can write down or talk yourself through a proof of why a program will halt, you have solved a computable problem. By definition. When you say "we can decide a whole lot of programs halt", you're talking about plain Jane computation.
If you have knowledge that any program that I give you will or will not halt, then you're a hypercomputing oracle for the halting problem, and you can do magic, basically.
There's a world of difference between the two situations. And I'm fairly sure (but not positive) that there's no way for you to prove in finite space and time that you are an oracle for the halting problem.
Edit: for more on that last bit, see https://pdfs.semanticscholar.org/19f9/0c34cdda43efdcf0831b2f...
For example, Gödel's incompleteness theorem is a technical result stating that certain definitions of "model theoretic truth" in classical set theory are incompatible with certain other notions of "truth as provability". Both the statement and its consequences have been so massively oversold for most of the past century that you should never use it in a discussion - think of it as a logical version of Godwin's law.
Similarly, no free lunch type theorems are formally the same as the statement that you cannot compress all n byte sequences into less than n bytes, which is really obvious for cardinality reasons. Again, there is no magic, just clever reductions.
The argument behind the halting problem also applies to your brain and anything that is somehow an abstract model of computation. More fundamentally, the fact that the self halting problem is undecidable is simply an instance of Cantor's theorem, it's not something that can be avoided.
Mathematical logic and therefore computers can be used to talk about infinity and more. Logical "paradoxes" are not a problem either. Some may be genuine proofs of inconsistency of some logical theories and others are simply theorems. The usage of the word "paradox" in natural language is simply imprecise.
---
I could go on, but really I don't take offense at any particular point. What bothers me is that you seem to be overselling mathematical results to argue a non-mathematical point... If you really want to apply, e.g., Gödel type theorems to discussions about your brain, you would first have to argue that the assumptions of Gödel's theorems apply. For instance, you could argue that the definition of model based truth in Peano arithmetic is something that your brain can decide. Then it would follow that what your brain does is uncomputable.