6 ms·
About AlphaZero particularly, a few things must be kept in mind. First, AlphaZero still makes use of a Monte Carlo Tree Search algorithm to search for good mov
by YeGoblynQueenne 9y ago
About AlphaZero particularly, a few things must be kept in mind.
First, AlphaZero still makes use of a Monte Carlo Tree Search algorithm to
search for good moves. MCTS is a powerful algorithm with a very limited scope:
zero-sum, perfect information games. So for instance, it would be very
difficult to see how to use MCTS-based AlphaZero in, e.g., training
self-driving cars.
Second, the AlphaZero architecture is precisely mapped onto a checkerboard and
will not learn anything about games that don't use a checkerboard, or any
situation that is not possible to model as a game played on a checkerboard.
Third, the AlphaZero architecture is also precisely mapped onto the range of
moves of pieces in chess, shoggi and go. Again, AlphaZero would be useless in
any game that used pieces with different moves (e.g. a piece with a zig-zag
move, or a piece allowed to move in spirals etc).
All of the above of course can be mitigated with different architectural
choices, but to make those choices, implement them and validate them will take
a great deal of time.
So, AlphaZero doesn't mean we're closer to _general_ AI.
Quite the contrary: it's a very specialised form of AI that will be very
difficult to use in any different task than chess, shoggi or go.
- pas 9y agoHow does AlphaZero handle the Knight in chess?
- YeGoblynQueenne 9y agoThere's a vector encoding "queen moves" and another encoding "knight moves". Between them, they cover all possible chess moves. Knight moves obviously are modelled by "knight moves". Edit: https://arxiv.org/pdf/1712.01815.pdf https://arxiv.org/pdf/1712.01815.pdf The information you want is on page 13, last paragraph. The previous paragraph describes the representation of the boardd.
- simonh 9y agoThose vectors are only used to generate the move trees though. That part of the architecture is common to pretty much all MCTS board game AIs ever. The value in AlphaZero is in the neural nets used for the expert policy and the value functions and those don’t have anything about the game rules encoded into them at all. I agree it’s probably quite constrained in the range of possible applications. Everyone was expecting Deep Blue to revolutionise AI applications too. I know the tech is different, but the fact it seems optimised for a highly constrained problem domain isn’t, and in fact arguably the problem domain addressed by deep blue seemed for a long time to be much more general. How adaptable is AlphaZero to arbitrarily multidimensional grids though?
- YeGoblynQueenne 9y ago>> The value in AlphaZero is in the neural nets used for the expert policy and the value functions and those don’t have anything about the game rules encoded into them at all. They do, in the form of their inputs that are basically vector representations of a checkerboard. It's obvious that the two (types of) networks can learn something useful from that particular representation of a problem. But- other representations, of different problems? That is not obvious.
- abecedarius 9y ago> Everyone was expecting Deep Blue to revolutionise AI applications too. That doesn't match my memory at all. The reaction then was dominated by the likes of "this is super-narrow, not real intelligence". (The 80s did have a lot of hyped expectations of related tech, it's true, but that was around 10-15 years earlier.)
- sgt101 9y agoI thought Watson would; what's happened to that project is a salient lesson.
- AndrewKemendo 9y agoSo, AlphaZero doesn't mean we're closer to _general_ AI. Quite the contrary: it's a very specialised form of AI that will be very difficult to use in any different task than chess, shoggi or go. This is a very true statement and one that I think a lot of people who aren't in ML/DL, but are "worried" about AGI, miss. There is however a common thread with everyone in AI, that they tend to think of AGI as "One algorithm to rule them all." As a practitioner and AGI researcher however I think that AGI is more of a system of specialized or narrow AI tasks that can together solve all tasks. At the risk of oversimplifying and anthropomorphizing, this type of problem solving is functionally how we do it as humans. So having a corpus of solved narrow systems (discrete known rule space in the sense of AlphaGo etc...) that is "activated" by an executive function which can recognize the problem set and then pass subsets of a larger problem to the narrow solutions. Those solutions are then "backpropagated" and synthesized into the general problem solution. In that sense, I would argue that narrow solutions like AlphaGo etc... do get us closer to General AI because they grow the corpus of solution paths for the general problems.
- s1dechnl 9y agoI disagree. The naming convention [Artificial Intelligence] still is a large shoe that these purpose built applied engineering solutions have yet to fill. Meanwhile, for profit/notoriety/marketing people want to trample on yet another name space? What you just described is essentially the architecture of a self-driving car. It's yet another applied engineering solution of Artificial Intelligence. It is not Artificial General Intelligence. Scaling/Distributing the computational space of an applied Artificial Intelligence solution is not Artificial General Intelligence. This is the same thing that lead to optimization algorithms being called Artificial Intelligence. If you aren't able to maintain foundational distinguishment, you lose track of what you're searching for and trying to achieve. Outwardly, you capture more money and attention. Inwardly, you become unraveled and lose your capability to solve the elusive problem. Eventually after much fame, wealth, and feigned 'success', one asks themselves : Was it worth it? Depends on what your original aim was.
- AndrewKemendo 9y ago
- s1dechnl 9y agoI would just like to echo this comment as it exactly and indefensibly strikes out Weak-AI from Artificial General Intelligence. I fail to see how specially crafted optimization algorithms continue to receive higher attribution. First, they were renamed to Artificial Intelligence. A great deal of cash exchanged hands based on this attribution. To retain the spirit of the true definition of Artificial Intelligence, the naming convention : Artificial General Intelligence was crafted. Even then, the allure of cash/fame progresses to taint that whereby people are calling expositions of Optimization/search algorithms Artificial General Intelligence. It saddens me that mankind has progressed the ubiquitous and beautiful information age into the disinformation for profit age from top to bottom and bottom to top.
- laughingman2 9y agoIt's highly disingenuous to portray the researchers in machine learning as some sort of sell outs . Hinton was working in neural nets even in the time when they were uncool and fringe area of research. They didn't sell out AGI for some profit motive . The thing is we don't have any clear path laid to follow to achieve AGI . Various paths proposed earlier turned out over optimistic and dead end. The current "weak AI" you disparage is an achievement expanding generalization of problem spaces using unified methods way better than past systems. So even if our current progress seems disappointing , please examine what's happened and why research has taken the current path before tarnishing researchers with unjustified quips.
- s1dechnl 9y agoI don't recall portraying anyone in that light. I highlighted the nature of an industry. Who aligns themselves to this nature was left unstated. If this clear and present truth offends, its likely because it has merit and is validated by the level of offense one encounters. It quite clearly cannot be defended. If you believe what I have stated can be, you're more than welcome to produce sound arguments that try. We can walk through an incredible amount of examples of what I have stated together. > Hinton was working in neural nets even in the time when they were uncool and fringe area of research. They didn't sell out AGI for some profit motive . Listen to what you're saying.... > neural nets even in the time when they were uncool and fringe area of research. And yet, it has made others billions and continues to mint money. Whose centered on the fundamental problem as opposed to conducting applied engineering for profit? Whose over marketing themselves and their efforts as fundamental theoretical research when its more or less optimizations for applied engineering? What ideation is new and what is simply relabeling old pioneer's work as one's own? Who minted LSTM? Whose name remains all over it? Who made a mockery of a prominent contributor? Who no longer discloses the details of their work given the proven nature of the industry? If the critique doesn't apply let it fly. Btw, Hinton recently stated : https://www.axios.com/artificial-intelligence-pioneer-says-we-need-to-start-over-1513305524-f619efbd-9db0-4947-a9b2-7a4c310a28fe.html https://www.axios.com/artificial-intelligence-pioneer-says-w... Whose taking this advice? Whose funding people thinking outside of the box? Whose hiring someone whose thinking outside of the box? So, whose really and whole-fully centered on solving AGI for the sake of solving it? It's fun to market yourself as doing so for increase prominence/money. It's a whole other ball game to be internally oriented and structured in pursuit of it. Applied engineering/optimizations of Weak AI for business applications is not AGI research. > The thing is we don't have any clear path laid to follow to achieve AGI The path was always there. One only need pursue it for pursuits sake. No exits. No distractions. No business case. No payouts. A desk pen/paper. > Various paths proposed earlier turned out over optimistic and dead end. False. Various paths proposed earlier are the same ones being turned into profitable solutions today. They are the same ones underlying the bulk of white papers today. There weren't over optimistic. They weren't dead ends which is why various groups are using them to suck down billions of dollars today. What happened that caused the previous AI winter is that people rushed fundamental research into applied engineering. The same as what's happening today and will lead to a 'winter' for various groups who are deep in it. Statistics overshadowed fundamental mathematics and understanding. Brute force over-took intelligence... And yes that leads to dead-ends .. clearly. Yet, here we are again. A bunch of statistical brute force machines being over-marketed for profit leading fundamental research into a dead end. Google has spoke out about it, Hinton has, Microsoft has, yan lecun to echo my sentiments. Yet, the allure of profit/notoriety continues to attract the same thinking/soon to be failed approaches to something that is fundamentally beyond current work. > The current "weak AI" you disparage is an achievement expanding generalization of problem spaces using unified methods way better than past systems. It's distributed statistical optimization. There's no need to fluff me. I am centered on it. It works via large data sets, limited and specific state spaces, and large amounts of computational resources running tons of iterative steps. It is brittle and not general. It works today because we have enough computing resources such that we can brute force various problems. Problems that are privy to statistical patterning. > So even if our current progress seems disappointing I never stated I was disappointed nor did I disparage anyone. I stated more clearly the nature of the progress, the goals and driving forces, and the fast approaching limitations. This frees one to respectfully acknowledge and consider what's been done and move forward and beyond it to more capable pastures. > please examine what's happened and why research has taken the current path before tarnishing researchers with unjustified quips. Please reduce your sensitivity level so that you can reason yourself to higher planes of consideration. Less emotion and more reasoned/truthful admission. I know exactly what happened to research in a time's past which informed me as to how to conduct myself in the present. I know exactly what driving forces are and I know quite clearly as we all do as to how they influence people. Hinton as well as others now public agree. Microsoft/Google have stated much of the work is overhyped. Hinton says pioneers should scrap everything and start over. LeCun stated that deep learning is approaching a brick wall and overhyped. It's exactly what I said to myself years ago and am stating now and there's zero problems with stating this truth as so many have no brought themselves to state.
- yonkshi 9y agoI don't agree with this sentiment, although I agree that AI is not nearly at the level of the hype that pop culture makes it out to be. AlphaZero is still a significant contribution to 'AGI' that shouldn't be buried. It's true that AlphaZero's knowledge is unable to be generalized for other systems, but its biggest contribution is a _stable_ RL system that can solve problems that no other systems can. This is the first piece of the puzzle of more general AI. Generalization, I would consider as the second piece of the puzzle to more general AI. Generalization may be achieved with potential research in transfer learning, model based RL, symbolic network. But without a stable RL algorithm such as DQN as foundation, generalization has nothing to stand on.
- s1dechnl 9y agoHaving not defined what intelligence is and having not declared the nature of General Intelligence, you're sure an aspect of clearly defined weak AI is a significant contribution to 'AGI'... Interesting. > It's true that AlphaZero's knowledge is unable to be generalized for other systems Interesting admission. > This is the first piece of the puzzle of more general AI. The first piece is generalized intelligence. Architecturally, it looks nothing like Alpha Zero. However, you feel : > Generalization, I would consider as the second piece of the puzzle to more general AI. How is that the second piece? It's the piece. > Generalization may be achieved with potential research in transfer learning, model based RL, symbolic network. But without a stable RL algorithm such as DQN as foundation, generalization has nothing to stand on. So you're of the belief that current approaches are compatible with and are the underpinning of Artificial General Intelligence while Hinton is convinced one needs to scrap it and start over. Sound advice is being ignored and there is a clearly entrenched decision to continue pushing along w/ iterating weak AI. I came here to test the waters.. The commentary and the K-value feedback I've received so far informs me quite profitably.
- yonkshi 9y agoHinton's criticism is very valid, but it's not quite about AlphaGo and its branch of ML. his criticism revolves around supervised learning and back prop that cannot be used to achieve the so called AGI. Because ANN is nothing like our brain's real NN. When Hinton gave the speech back in 2014, NN had a huge explosion of hype, and NN was mostly for supervised learning, which is really only good at classification and regression problems, it cannot make decisions outside of its training. The famous DeepMind DQN paper (the core of AlphaGo) were published after Hinton's talk. the DQN paper practically opened a new chapter in reinforcement learning field. I am not sure if you are familiar reinforcement learning. RL is learning by trial and error, model-less and largely non-bayesian, similar to how humans learn. Up until AlphaGo, RL field was stuck in a limbo because it was having a very hard time learning non-linear problems (which is the majority of problems in nature) When I say generalization, I meant generalization of knowledge. Generalization is the second piece of the puzzle because, even as humans, we learn from experienc. After enough examples we began to generalize. Up until DQN came out, we couldn't even effectively learn. It's the equivalent of a human baby with severe memory problem. With deepmind's DQN, we can achieve much more stable learning on non-linear systems, and we can begin to add components such as generalization (such as transfer learning), intuitions (such as intuitive physics), symbolic network. I am not too sure what you meant by weak AI and general AI. For me, an AI which can learn similar to how humans learn, able to use apply generalized knowledge when facing a brand new problem, independently think and make decisions without human assistance, that's general enough. Yes, much work needs to be done, but I don't believe this is the wrong direction we are going. Though I'd be glad to be proven wrong and I am very fascinated by this debate, if you would like, we could continue discussing it over email/chat?