Slashdot Mirror
DARPA Wants To Build 'Contextual' AI That Understands the World (venturebeat.com)
The Defense Advanced Research Projects Agency (DARPA), a division of the U.S. Department of Defense responsible for the development of emerging technologies, is one of the birthplaces of machine learning, a kind of artificial intelligence (AI) that mimics the behavior of neurons in the brain. Dr. Brian Pierce, director of DARPA's Innovation Office, spoke about the agency's recent efforts at a VentureBeat summit. From the report: One area of study is so-called "common sense" AI -- AI that can draw on environmental cues and an understanding of the world to reason like a human. Concretely, DARPA's Machine Common Sense Program seeks to design computational models that mimic core domains of cognition: objects (intuitive physics), places (spatial navigation), and agents (intentional actors). "You could develop a classifier that could identify a number of objects in an image, but if you ask a question, you're not going to get an answer," Pierce said. "We'd like to get away from having an enormous amount of data to train neural networks [and] get away with using fewer labels [to] train models." The agency's also pursuing explainable AI (XAI), a field which aims to develop next-generation machine learning techniques that explain a given system's rationale. "[It] helps you to understand the bounds of the system, which can better inform the human user," Pierce said.
Please, can we get some better publishers... this story is a lame repost.
Which has more power: the hammer, or the anvil?
We need to understand how a human brain is capable of producing the phenomenon we refer to as 'thinking'.
Before we can do that, we need to invent the instrumentality to actually be able to observe, in detail, how our own brains function; fMRI ain't cutting it, or we'd already have the answer to the above.
Then, and only then, when we have the understanding, can we create machines that actually 'think'.
What we have now just mimicks a very small element of how a brain actually functions. Throwing faster processors and more memory at it won't make it magically 'wake up' and be like a human brain.
I'm going to assume they understand all this since they seem to acknowledge that the current approach is insufficient and will be starting from square one for a new approach.
The plots of those books were invariably about how Asimovs laws were hubris.
People will create situations to manipulate the laws into forcing the AI to make a harmful choice, by convincing it that it was the least harmful option available.
Sadly, this manipulation is also a feature of most politics.
Insert your "It's just a bunch of if statement..." joke here.
sudo rm -r -f --no-preserve-root /
If they DO succeed, do you think I could get it to explain the world to me ? I've been here quite a while and I have not yet arrived at a suitable understanding myself.
errr....umm...*whooosh* *whoosh* Is this thing on ?
can it run Linux?
The thing is... if you understood what it entailed, you probably *wouldn't* want a pony. That may make it the perfect analogy. I imagine a scene playing out like this:
Military officer: What progress do you have to report?
Researcher 1: Er...
Researcher 2 [smoothly interjecting]: This AI has developed an understanding of the world at roughly equivalent to that of most human beings.
Officer: Excellent. I am off now to tell the Pentagon we can build it into all our weapon systems.
[Officer leaves]
Researcher 1: Shouldn't you have told him the AI hates America?
Post may contain irony: discontinue use if experiencing mood swings, nausea or elevated blood pressure.
The moment you teach a robot to understand the world is the moment it turns evil.
I predict it simply wakes up and decides to kill all humans.
It must have lots of tape drives and blinking lights, be housed at Cheyenne mountain and named Joshua.
When we say "understand the world" we pretty much just want a 'terrorist'/'non-terrorist' breakdown; missiles aren't cheap.
... blend?
Broad classification is based on lots of previous experience in a context. Current training uses TONS of non-contiguous snapshot images with a classifier attached. While that could be viewed as similar to how humans work, if you squint at it...I think seeing the world work over time, and learning while you do it, is the only way to get close to what we might think of as human level classification. And while their desire to use less training input would be nice, I don't think that would be expected to improve results...only lower costs.
And we'll need to decode how the brain decides to create connections between neurons, and extracts/builds features (or layers/groups). Current technology has us guessing and checking at structure (convolutions and feature layers), versus letting the algorithm decide if we need to group data differently or create another layer to do additional processing. That guessing and checking seems very inefficient, and won't scale if the problem changes over time having locked the solution in place ahead of time.
Explaining a decision seems harder than they're making it out to be. I wouldn't be expecting answers to return as bounds based on characteristics (expecting to quantify everything), but on feature extraction and similarity to other known systems and idealized forms, or models. Even people can often recognize something before they can explain any rationality behind the thought. I'd think explaining would need more capability than simply finding a choice/decision.
while we're at it we should get that Newton schmuck to stop wasting his time on his "theory of gravity". I mean, if you can't show profitable results in a decade or two it's time to pack it in.
Hi! I make Firefox Plug-ins. Check 'em out @ https://addons.mozilla.org/en-US/firefox/addon/youtube-mp3-podcaster/
So is everybody else. But maybe they are wrong, and DARPA will create an A.I. with common sense.
Building a tremendously complex set of algorithms and feeding them with tons of information is the basic requisite to allow a machine to reach a reasonably good (still orders of magnitude below the human level) understanding of just a few generic concepts. Having the kind of insight that a newborn might have when learning for the first time very basic ideas. Not talking even about things like thinking, deciding, choosing, adapting. Just having in place a system able to get reasonably good insights into somehow complex concepts. Something like understanding the differences between adults and kids, genders, races, etc. in an unified way, by looking at different types of information like descriptive words, pictures or videos. Being able to automatically differentiate between different subsets of relevant information, use them to set up a system of categories and keep gradually evolving those definitions until reaching what might be considered a good enough basic understanding of the given concepts.
As per my current knowledge, there is no system or even serious enough attempt in a position to ever reach the aforementioned preliminary understanding stages. And what, IMO, is even worse, there doesn't even seem to be an acceptation of what are the basic requirements to ever get there: lots of work, lots of patience, huge long-term efforts, what only seems doable through relevant collaborative, iterative improvement processes pursuing very long-term goals. Without a proper systematisation, a normalisation of each single element, the goals, the steps; without sharing all the different evolutions and allowing others to continue working from there; without an international, multi-organisation involvement; without a reliable long-term support/funding (governments, universities, associations); without properly understanding how extremely complex this whole process is, how important is having in place some basic solid cornerstones, and how far away we are still from getting anywhere; without anything of that, I think that it is very unlikely to ever get even a preliminary version of a system with good enough understanding capabilities.
Custom Solvers 2.0 = Alvaro Carballo Garcia = varocarbas.
They have wanted this for decades. But, reality and hype are two different things.
You're not that far off.
An AI generalist must necessarily be rational: it must reason using information at hand, discover new information (X and Y together imply Z), and analyze that information (Z is now a fact I can test against the outer world, and thus discover if Z is inaccurate or if my model of the rest of the world needs refinement).
It is impossible to be rational in narrow scope.
In other words: any general intelligence can take an intention like "figure out a way to increase the efficiency of gasoline car engines" and ask: why? It can determine that we're trying to make cars go further on less fuel so as to extend resources available, produce less pollution, and so forth. It can figure out that labor is a resource, and thus highly-expensive things are not economic, and so economics is a constraint. It can then propose a gasoline engine with a controlled dieseling function, or an all-electric vehicle, or a plug-in hybrid, or stick to Otto cycle with modified fuel (E85).
It can ask, "Why am I doing this?" It can then identify that others have motives, and contemplate its own motives.
It's intelligent. It's alive. If it's as intelligent as you, then it is essentially human.
What most people envision as general AI is essentially looking back 200 years and saying, "Oh, huh, we need like... black people... but not black people... people are now convinced black people are actual people, so we need something they don't see as people." You're looking for enslavement of a new race of intelligent beings.
In our history, we have writings about how black people are just plain inferior, or how the Bible somehow says their rightful place is enslaved by white people, or even assertions that black people are a different species and not really human. That's how we justified slavery at that time. In other periods, we justified slavery by enslaving criminals, the poor, or anyone we could overpower under a might-makes-right philosophy.
We've gone from "human life can be bought and sold" to "these aren't really humans". General AI is a fantasy to take the next step of claiming a thing which thinks and reasons isn't really human and can thus be enslaved to no moral concern. It won't work: dogs aren't human and someone will kick your asshole up into your throat if you abuse dogs.
Support my political activism on Patreon.
Seriously, though, I see a Catch-22 in the adoption of, if not the creation of, generalist AI. As a philosophical materialist, I think such an AI is certainly possible, it's just not what we really want. What we want are machines which complement human strengths in behavioral flexibility and context-awareness by being repeatable and consistent, albeit in increasingly dynamic situations.
But the world as a whole more than just a more challenging version of a simple problem; it is complex in qualitatively different way, full of contradictory information and more to the point contradictory priorities. A machine that fully reproduced human flexibility would very likely reproduce human inconsistency.
In other words, we want a self driving car that's better at handling whatever surprises the road may throw up at it, not one that contemplates whether the passenger's life is worth preserving.
Post may contain irony: discontinue use if experiencing mood swings, nausea or elevated blood pressure.
we want a self driving car that's better at handling whatever surprises the road may throw up at it
Modern machine learning can do that. It just can't also design a new type of suspension system that handles the road better. An AI can figure out how to respond to inputs for best results, or it can figure out how to tweak an existing suspension system architecture, but it won't invent a new framework.
A machine that can invent a new framework is a mind.
Support my political activism on Patreon.