Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

The Difference Between AI, Machine Learning, and Deep Learning

blogs.nvidia.com

Post reply on HN
Print view View original

71–80 of 107 posts

Re: The Difference Between AI, Machine Learning, and Deep Learning

#71

Post by eli_gottlieb »

cgearhart wrote: post #61
ktta wrote: post #12

Also, an interesting read related to this would 'the AI effect'[0]. A lot of the stuff Deep Learning/Machine Learning is able to do today would be looked at as something that only 'true' AI (whatever consensus on what that means is; I think of it as AGI) would be able to do. But as soon as we are able to solve a problem that we think (feel?) only true AI (AGI) would be able to solve, as soon as we know how it was sol…

I would argue that there is actually an unrecognized inverse of this effect, too. People have a tendency to conflate domain-specific human-level performance with "intelligence" (for some definition of that word). Then suddenly we jump to "DeepMind won at go, therefore humanity is on the brink of creating Skynet." Deep learning has been like this so far, because it works in many new domains that have resisted previous…

>"Hard" problems are multi-optimization, where even the definitions aren't very clear: "learn as much as possible and live a happy life while making a productive living for yourself". Turn that into an objective function...

Minimize the informational free-energy of the product of the multiplicatively inverted brain-to-body reinforcement-learned energy function of reward with the brain-to-body reinforcement-learned energy function of punishment, via active inference?

Beh, those words are messy. The point is to talk about a Gibbs distribution with energy functions Reward(X, Y) and Punishment(X, Y), such that the "total" energy function is E(X, Y) \propto Punishment(X, Y) - Reward(X, Y). This then gives us a "goal distribution" for an active inference agent (like a human), defined "up to" the reinforcement-learned energy functions, whose limiting functions (need a lot of functional analysis and probabilistic reasoning over function spaces, there) are the "ground truth" causal relations by which the world causes reward signals through the body and its senses.

https://arxiv.org/pdf/1503.04187.pdf

Re: The Difference Between AI, Machine Learning, and Deep Learning

#72

Post by eli_gottlieb »

svantana wrote: post #16
argonaut wrote: post #13

Earlier quoted context omitted.

I hear this repeated a lot. But I've never actually seen a source from, say, the 90's or the 2000s, actually saying something like "if we're able to classify images better than humans, that's true AI." I suspect it's just hearsay that's been repeated over and over again.

There used to be a concept of "AI completeness" which meant that some problems could only be solved by True AI, which would also be able to solve all other human-level problems. Doug Hofstadter writes in Gödel-Escher-Bach that he believes grandmaster-level chess is AI complete. That was obviously false in retrospect, and it's becoming increasingly obvious that there is no AI completeness at all.

>it's becoming increasingly obvious that there is no AI completeness at all.

Bingo! The "general" part is the ability to learn new structures and task structures from environmental cues, and then construct informed prior beliefs about those new tasks and structures using causal relations to previously-observed tasks and structures. Nothing more!

Re: The Difference Between AI, Machine Learning, and Deep Learning

#73

Post by eli_gottlieb »

reacweb wrote: post #20
argonaut wrote: post #13

Earlier quoted context omitted.

I hear this repeated a lot. But I've never actually seen a source from, say, the 90's or the 2000s, actually saying something like "if we're able to classify images better than humans, that's true AI." I suspect it's just hearsay that's been repeated over and over again.

In the 90's, John McCarthy was calling Go "the new Drosophila of AI". Now, we have AlphaGo that is "just" a bunch of "generic" deep learning with MC. Maybe John McCarthy was right and the ideas behind AlphaGo are the first steps toward true AI. I am very impressed.

Sounds like McCarthy was right: the point of Drosophila is that it's a very small, simple model organism. You're not supposed to "solve biology" by modelling something in Drosophila, you're just supposed to be able to narrow the field to hypotheses that are not so flagrantly stupid they fail in a fruit-fly.

Re: The Difference Between AI, Machine Learning, and Deep Learning

#74

Post by eli_gottlieb »

visarga wrote: post #58
denzil_correa wrote: post #34

Earlier quoted context omitted.

Basically, Deep Learning is now back on the map due to two things - (1) Large amounts of data available and (2) Faster processing power viz. GPUs.

Well, yes, mostly, but there also have been genuine discoveries in the last 10 years. We can now train deep networks because we learned how to regularize - before it was impossible because of vanishing gradients. We can have even 1000-layer deep nets, which would have been unthinkable. Also, there are some interesting approaches to unsupervised learning like GANs and VAEs. We learned how to embed words and not only w…

>We can now train deep networks because we learned how to regularize - before it was impossible because of vanishing gradients.

Those are two different things. Vanishing gradient problems were ameliorated by switching from sigmoidal activation functions to rectified linear units or tanh activations, and also by dramatically reducing the amount of edges through which gradients propagate. The latter was accomplished through massive regularization to reduce the size of the parameter spaces: convolutional layers and dropout.

Stuff like residuals are still being invented in order to further banish gradient instability.

Re: The Difference Between AI, Machine Learning, and Deep Learning

#75

Post by j1vms »

ktta wrote: post #12

Also, an interesting read related to this would 'the AI effect'[0]. A lot of the stuff Deep Learning/Machine Learning is able to do today would be looked at as something that only 'true' AI (whatever consensus on what that means is; I think of it as AGI) would be able to do. But as soon as we are able to solve a problem that we think (feel?) only true AI (AGI) would be able to solve, as soon as we know how it was sol…

And on a related note, "better than humans", "comparable to human intelligence", and phrases like these since the early days of AI have always been a red herring. Not the least because these phrases do not imply what we are really comparing - better than all humans? some humans? How about just a person? What about just better than a 4-year-old?

The criteria should have always been "faster / more efficiently", "more accurately", or "both" in comparing AI to human intelligence for a given task or activity. By this measure, AI had enduring production-ready successes since at least the 1980s. Let's say that was software that ran on post-VLSI hardware (1+ million transistors). Now, we have deep learning deployed on multiple GPUs each with a 15+ billion transistor count.

True AI is starting to look philosophically like a mirage, not the least because we may not ever have a strict meaningful definition for intelligence. What we consider intelligence varies, evolves, changes shape, consistency and predictability even if we consider just one of the various people we interact with on a given day.

As one of the latest, I think Brian Eno (https://news.ycombinator.com/item?id=12027055), like others before him who have formalized a similar approach to the matter of intelligence, is on the right track. This approach to deciding what constitutes True AI would do much to counter and eventually prevent the "AI effect".

Re: The Difference Between AI, Machine Learning, and Deep Learning

#76

Post by Cybiote »

visarga wrote: post #48
seanwilson wrote: post #18

Earlier quoted context omitted.

> But as soon as we are able to solve a problem that we think (feel?) only true AI (AGI) would be able to solve, as soon as we know how it was solved, it is no longer a mystery that warrants amazement and we argue that it is not real intelligence, just like the link below states. I think the issue is that current examples of AI can mostly only solve problems in a single domain instead of in a variety of domains. Deve…

> once we have a comprehensive AI program that can play many kinds of games Before AlphaGo, DeepMind released a reinforcement learning algorithm that could play many Atari games just from the raw pixels on the screen, in many games surpassing humans. The same algorithm. https://arxiv.org/pdf/1312.5602.pdf Reinforcement learning is a general framework for learning behavior from acting in an environment with the purpos…

> Atari games just from the raw pixels on the screen

It's important to distinguish between what sorts of games work well under this method and what sorts do not. Games that are variations of pole balancing, like Pong, fare better than more complex games like Asteroids, Frostbite or Montezuma's Revenge.

> Saying that AlphaGo is limited because it only knows to play one game, is like saying that humans are limited because Lee Sedol could only master at world level one game.

It's nothing of the sort. AlphaGo is a machine in the Turing sense. The neural network is a program that is the result of a search for a function specialized to playing Go. This machine, the program that the parameters across the edges in the graph represent, is logically unable to run any other program. Lee Sedol is a Universal Machine in the Turing sense, any statement contradicting this makes no mathematical sense.

> We limit software to specific domains only on account of efficiency, not because algorithms are fundamentally limited.

It is well known within the literature that these models do not make best available use of information when learning. They are exceedingly inefficient in their incorporation of new information. Issues include improper adjustment of learning rates, not using side information to constrain computation, having to experience many rewards before action distributions are adjusted in the case of reinforcement learning, samples per example in supervised learning. Note that animals are able to learn without explicit labels and clear 0/1 losses.

Humans and animals generally, even in the supervised regime, are vastly more flexible in the format the supervision can take.

For an example, look into the research on how children are able to generalize from ambiguous explanations as "that is a dog" and why difficulty in learning color from this kind of "supervision" shows just what priors are being leveraged to get that kind of learning power.

See here for an excellent overview of limitations in our current approaches to AI: https://arxiv.org/pdf/1604.00289v2.pdf

> "Learning without forgetting"

That's a great paper but it does this by minimizing prediction error drift by comparing before and post performance on the old task while learning the new. I do not know that this method will scale with increasing task numbers, considering Neural Networks are already difficult and energy-time consuming enough to train as is.

Re: The Difference Between AI, Machine Learning, and Deep Learning

#77

Post by ihaveahadron »

IMO: With so many nonsensical things being in the article, it still seems like we already have AI -- making organisms that grow into brains (humans). And this so called AI still seems like ever more complex machinery. If the computers are learning, I doubt its from a vacuum, its from things (language, etc.), created by humans. I guess my point here is that AI suggests that the computers are "alive" conscious "entities". I don't see it.

Re: The Difference Between AI, Machine Learning, and Deep Learning

#78

Post by cosmolev »

jules wrote: post #51
mjn wrote: post #46

Earlier quoted context omitted.

I disagree that neural networks are necessarily mysterious. In most cases they can be straightforwardly understood as brute-force function approximation (approximation here meaning mostly interpolation, in some cases capable of a limited degree of extrapolation). You take many samples of a function (either because you have a huge data set, or because you can actually sample with a simulator), and incrementally modify…

Of course, but compared to how chess algorithms work it remains far more mysterious how exactly that produces an english description of a grid of pixel values.

So they take grid of pixels and produce an array of characters which may be interpreted by some humans (aka bags of chemicals).

Re: The Difference Between AI, Machine Learning, and Deep Learning

#79

Post by antognini »

oh_sigh wrote: post #35
thallukrish wrote: post #21

If I need to show 1 million cat images to train a neural net to see cats, I wonder how a human brain can figure out cats of any kind just by seeing one or two. Is there something fundamental we are missing in going about building these deep learning stuff ?

"one-shot learning aims to learn information about object categories from one, or only a few, training images": https://en.wikipedia.org/wiki/One-shot_learning

There's also the more challenging problem of zero-shot learning, e.g.:

http://jmlr.org/proceedings/papers/v37/romera-paredes15.pdf

"Zero-shot learning consists in learning how to recognise new concepts by just having a description of them."

Re: The Difference Between AI, Machine Learning, and Deep Learning

#80

Post by seanwilson »

visarga wrote: post #48
seanwilson wrote: post #18

Earlier quoted context omitted.

> But as soon as we are able to solve a problem that we think (feel?) only true AI (AGI) would be able to solve, as soon as we know how it was solved, it is no longer a mystery that warrants amazement and we argue that it is not real intelligence, just like the link below states. I think the issue is that current examples of AI can mostly only solve problems in a single domain instead of in a variety of domains. Deve…

> once we have a comprehensive AI program that can play many kinds of games Before AlphaGo, DeepMind released a reinforcement learning algorithm that could play many Atari games just from the raw pixels on the screen, in many games surpassing humans. The same algorithm. https://arxiv.org/pdf/1312.5602.pdf Reinforcement learning is a general framework for learning behavior from acting in an environment with the purpos…

> Reinforcement learning is a general framework for learning behavior from acting in an environment with the purpose to maximize a reward. It can be used, and was used, in multiple domains. AlphaGo used RL as well.

> Also, regarding other tasks: a neural net that recognizes cats can be easily made to recognize dogs too. A program that translates English to French can be made to translate other languages too.

I'm aware of these things. I was commenting on the shifting of the goal posts of what people call intelligent. In my opinion, when someone can deliver a concrete implementation of a computer that is competent (not even expert level) of many varied domains, then most people would call that intelligent. Right now, examples of AI are genius level at a niche domain and unable to do anything in all others (without tuning at least).

I don't think anyone would call an algorithm that in theory can apply to lots of domains intelligent. You need a concrete implementation to demonstrate this.

Post reply on HN