Can a Robot Learn a Language the Way a Child Does? (zdnet.com)

MIT researchers have devised a way to train semantic parsers by mimicking the way a child learns language. "The system observes captioned videos and associates the words with recorded actions and objects," ZDNet reports, citing the paper presented this week. "It could make it easier to train parsers, and it could potentially improve human interactions with robots." From the report: To train their parser, the researchers combined a semantic parser with a computer vision component trained in object, human and activity recognition in video. Next, they compiled a dataset of about 400 videos depicting people carrying out actions such as picking up an object or walking toward an object. Participants on the crowdsourcing platform Mechanical Turk to wrote 1,200 captions for those videos, 840 of which were set aside for training and tuning. The rest were used for testing. By associating the words with the actions and objects in a video, the parser learns how sentences are structured. With that training, it can accurately predict the meaning of a sentence without a video.

Comments

No

[Anonymous Coward]

No

Re: No

[Anonymous Coward]

Beat me to it. No. That would require understanding something that we actually know laughably little about. Their efforts would be better spent realizing software is just software and focusing on what software is and is not capable of, because software is not going to cease being software. Human beings are not machines running an OS.

Re:No

[sittingnut]

too bad zdnet editors, and then /. editors in summery, failed to question the unwarranted assumption in the article that we know how a child learns, instead allowed, and even amplified, the reporter's uncritical parroting of the "researchers" pitch.

Re:No

[hey!]

What if we rephrased the question, e.g., "What would an AI need to be able to acquire grammar and semantics by being trained on natural language sentences (the way human children are)?"

Those of us who have a mechanistic position on consciousness and intelligence see no theoretical obstacle to building a machine that does anything or indeed everything humans do. But many of us are dubious that AI will ever achieve true parity with the full range of human abilities. My doubts are economic in nature. I doubt that any such generalist AI will ever be the cheapest way to get whatever it is we want out of a machine.

Take the "AI" that's hot in the market now. It's not an AI like the robots in Asimov's storeis -- a mechanistic simulation of what people can do. The machine learning stuff being flogged by companies today is just a way of replacing people on certain tasks with something that is cheaper and in some case more consistent, albeit less versatile.

There's one exception to the rule that a generalist AI isn't really what we want, and that's if we want to prove a non-material soul is unnecessary for explaining anything about humanity. And I doubt anyone really cares enough about such a demonstration to pay what it would take to do it convincingly.

Re:No

[ShanghaiBill]

Chomsky famously called it a "black box inside their heads."

Noam Chomsky was being a bit modest. He did more than anyone to figure out what is going on inside that black box, and what innate language learning ability children are born with, which is far more than the "tabla rasa" theory pushed by behaviorists. Chomsky learned that all human languages, even those invented by isolated groups of children, have nouns, verbs, adjectives, and adverbs. All of them have words for discussing hypotheticals, and situations separated in both time and place from the here-and-now.

Chomsky wasn't right about everything. He believed that all human languages were based on recursive grammars, and for decades this was thought to be correct. But recently it was reported that the Pirahã language does not allow recursive phrases.

Re: No

[phantomfive]

Chomsky wasn't wrong, the reports of lack of recursion were wrong. (Even if the reports were correct, it doesn't mean Chomsky was wrong, he merely said humans are capable of recursion, not that they use it all the time or need to use it)

Re:No

[gweihir]

What if we rephrased the question, e.g., "What would an AI need to be able to acquire grammar and semantics by being trained on natural language sentences (the way human children are)?"

Simple: Actual intelligence. "AI" is a marketing term. No "AI" these days and for a long time to come (possibly forever) has actual intelligence. All we have is mindless automation that is not capable of insight.

Re:No

[gweihir]

Nobody knows how intelligence, insight and consciousness works either. Not even the neuro-"scientists" that like so much to claim otherwise, but pathetically fail when put to any real test.

Re:No

[Anonymous Coward]

"What would an AI need to be able to acquire grammar and semantics by being trained on natural language sentences (the way human children are)?"

To actually be AI. Nothing we have is anywhere close. The media and tech companies like to toss around the term, but it's all marketing and cluelessness.

Re: No

[Zero__Kelvin]

Nobody but you is claiming the A in AI stands for "actual", but your faux expertise on this subject found all over in these comments make me wonder if you have any "actual intelligence."

Re: No

[Zero__Kelvin]

Of course we are machines running wetware. To believe otherwise shows a remarkable dearth of self-awareness.

Re: No

[Zero__Kelvin]

It's a shame that you assumed that we have to know that in order to acheive the goal.

Re:No

[SqueakyMouse]

Because nobody knows how a child learns language. Chomsky famously called it a "black box inside their heads."

This means if we had a robot learning the way a child does, we would not be able to verify it was learning in the same way. It doesn't mean it's impossible to independently arrive at the same solution. It just means we wouldn't know it was the same method but why would we care? If the inputs and outputs are the same and the machine is efficient, then problem solved.

Re: No

[Anonymous Coward]

Nice imaginary world you've got there. Too bad it doesn't correspond with reality, kid.

Re: No

[Anonymous Coward]

Yes, but there is a complexity issue, just the number of neurons in a human brain. We are very far from understanding how any complex thought arises in the brain and there is therefore a lot of BS in the fields studying it

Re:No

[Anonymous Coward]

No

It worked with Johnny 5.

Re: No

[Zero__Kelvin]

You suck at trolling.

Re: No

[dj245]

Chomsky wasn't wrong, the reports of lack of recursion were wrong. (Even if the reports were correct, it doesn't mean Chomsky was wrong, he merely said humans are capable of recursion, not that they use it all the time or need to use it)

The reports of recursion in the Pirahã language seem disputed. I suppose that's what happens when you have an extremely small field of experts studying a unique language spoken by around 250 people.

Chomsky's assertion is marked as "not in citation given" in the wikipedia article on the Pirahã language. I don't care either way but taken together with the dispute about recursion in the language, it's an interesting nerd kerfuffle to witness.

Re:No

[Anonymous Coward]

It's "tabula rasa" , know-it-all.

Re: No

[Anonymous Coward]

And you suck at thinking.

No

[Anonymous Coward]

Because nobody knows how a child learns language. Chomsky famously called it a "black box inside their heads."

BS

[110010001000]

MIT has been claiming this type of BS for decades. They haven't done anything. Literally they have been talking about this since the 1970s. Think about it: if it worked it would have been incorporated into something like Siri and be worth billions. But Siri is pathetic.

Re:BS

[Anonymous Coward]

MIT has been claiming this type of BS for decades. They haven't done anything. Literally they have been talking about this since the 1970s.

If a machine is to be able to learn a language the way a child does, that machine needs to be able to experience the sensation of something like PEEING , for example.

A child may say 'Pee Pee' because he or she feels pressure building inside the bladder, that there is a need to release that pressure, and 'pee' is the action in which that pressure gets to be released.

Now, can a machine go through the same process?

If yes, then, the answer to whether a machine can pick up a language the way a child does might be 'Yes'.

If a simple thing like 'Peeing' can't go through the same 'pressure building' / 'pressure releasing' process, then, the answer gonna be a definite 'NO' !

Re:BS

[hcs_$reboot]

that machine needs to be able to experience the sensation of something like PEEING

Actually, a machine without sensors / feelings might a lot during the learning process, and your comment makes sense. However not sure "peeing" is the most important feeling in that process.

Re:BS

[Anonymous Coward]

Yes, because people without sight and hearing can't learn to communicate because they don't have the same sensory experience as normal communicators, right? And people born with paralysis who can't feel their peepee or other problems have difficulty learning these concepts, right?

A child can't really learn about a butterfly coming from a caterpillar and the metamorphosis process because tehy can't experience it, right?

This is about the stupidest argument I've heard against machine learning a language in a long time. Seriously, how much of what you've learned have you experienced first hand? How many 'feelings' actually *helped* you to learn? Yes, sometimes interest (which can be considered a feeling) helps learning but only if the student is non-motivated (another feeling).

Without feelings, I actually think it would be easier to teach a machine to talk about things. It might be difficult to understand the pressures of needing to peepee, but even that could be taught.

Re:BS

[gweihir]

Indeed. But they are not the only ones pushing things that inspire humans but are technological bullshit. Just think of flying cars, quantum computers, and the whole endless list of persistent failures along the same lines. There are other projects that are very long time (fusion, self-driving cars, etc.), but the are making persistent and meaningful advances all along the path. This stuff here does not. It just goes from meaningless stunt to meaningless stunt, because they have nothing.

Re: BS

[Zero__Kelvin]

The answer is of course yes, a machine that occasionally has the need to release built up pressure already exists. Of course your point is absurd, but I also wanted to point out that if that was somehow some critical component to the design it would be a complete non-issue anyway.

Re: BS

[Zero__Kelvin]

I love how you list a number of things that naysayers have traditionally cited as pipe dreams that are now either realized or on the verge of being realized to try to make a case for something you don't understand in the least being impossible. Seriously ... Give it up. "There will never be computers in most homes! It's impossible and the idea is absurd!" - virtually everyone I told that this would happen circa 1982

Re: BS

[gweihir]

I am nor a naysayer. I am a technology expert. Not sure whether you are equipped to understand the difference though.

Re: BS

[110010001000]

None of those things are being realized. Quantum computers for example are a sham. You just believe marketing hype.

Re: BS

[Zero__Kelvin]

You are a clueless incompetent. Off you go now ...

Re: BS

[gweihir]

You are mistaken. You should not project yourself onto others....

Re: BS

[Zero__Kelvin]

You are an idiot, but this is not new information. You have been told so many, many times.

Re:BS

[Anonymous Coward]

Last one was Cog btw. https://en.wikipedia.org/wiki/Cog_(project)

that's retarded

[Anonymous Coward]

whoooo let's train a neural net (oooh! aaah!) to predict some vector that can only be compared to other outputs from neural nets

if we use words describing leaky abstractions, and do syllogistic logic on those words, we can tie ourselves in semantic knots enough to pretend to have something interesting

Re: that's retarded

[Zero__Kelvin]

It worked for your post!

It's all fun and games...

[Anonymous Coward]

Until Star Fleet Command transfers it to a science facility for further research.

Luckily Trump will never exhibit enough covfefe to have to worry about such a things.

Re: It's all fun and games...

[Zero__Kelvin]

It's Space Force you enemy of the people! Star Fleet had colored folk. Space Force will be white with a touch of Orange spray tan.

No & Yes

[hcs_$reboot]

Not currently, because no human was smart enough to find the right algorithm. Yes, when Google or the Chinese make the algorithm. One difference though: it'll learn much faster than a human child.

Re:No & Yes

[gweihir]

One difference though: it'll learn much faster than a human child.

There is absolutely no reason to expect that.

Re: No & Yes

[Zero__Kelvin]

Except of course a basic knowledge of human anatomy and electronics.

Re:No & Yes

[hcs_$reboot]

There is. It's hard to control the way a brain works, to teach it new things optimally, and it has many limits, and can hardly be acutely controlled from the inside. A human made algorithm on the other hand, once created, should be optimizable in a much easier way.

Re:No & Yes

[StikyPad]

There are lots of reasons to expect that, namely that we use computers precisely *because* they are faster (and more accurate) than humans at processing information.

Re: No & Yes

[gweihir]

That one would tell you that the robot would be several orders of magnitude slower. If you actually had that knowledge. Electronics is slow and large in comparison to neural tissue.

Re:No & Yes

[gweihir]

They are not. They are faster at doing simple transformations on digital information. In the analog space for difficult problems (and learning language certainly is a difficult problem) computers are somewhere between very slow and incapable.

Re: No & Yes

[Zero__Kelvin]

Wow, you are one stupid motherfucker

Re:No & Yes

[RespekMyAthorati]

That is because AI is at the same place that electronics was a century ago: a lot of interesting ideas, but little or no results.
Machines may be able to learn to communicate as children do - but not in my lifetime or yours.

No

[AHuxley]

See AI winter https://en.wikipedia.org/wiki/....

AIs are trained on grammatical sentences

[aberglas]

That is what "statisitcal" language learning is about. Particularly useful for translation.

Your recharacterization is indeed much better than "to learn the way children learn" as we do not know how children learn.

AIs can already do many things better than people. If it would ever be as intelligent as a small child on everything, it would be a lot more intelligent than an adult in many other things.

An AI is not a human. It is a different beast entirely.

Re:AIs are trained on grammatical sentences

[gweihir]

Indeed. The thing with automated translation is that it does not require an independent world-model. The world model and the placement of objects, attributes and actions in it is already contained in the input. That is the only reason machines can actually do translation better than by using a gigantic look-up table.

Re:AIs are trained on grammatical sentences

[Anonymous Coward]

There is a *LOT* more to machine translation than a giant look up table. There are far, far more combinations of words and phrases than can simply be looked up. You have to know which parts of which sentences correspond, and how to learn what a translation is telling you. Certainly, people cannot learn to speak a language just by studying translated texts. (Well, I suppose they can if they are really clever and work really hard, the Rosetta stone comes to mind.)

That said, having a large corpus of translated text is very helpful. UN documents are one such source.

Re:AIs are trained on grammatical sentences

[gweihir]

You actually have to pretty much "know" nothing for machine translations, as machines cannot do that. Machines can do rules (id precisely enough defined), they cannot do understanding. This is what will keep automated translations on a low quality level permanently. High quality translations will remain a domain of smart humans with a deep understanding of both source and target culture.

The funny thing is that the only situation where machine translations can go beyond that is actually the giant look-up table as that can (to a limited degree) encode the understanding of a human being. Step outside of that and the translation becomes bereft of any insight into what is being translated.

Re:AIs are trained on grammatical sentences

[Anonymous Coward]

The world model and the placement of objects, attributes and actions in it is already contained in the input.

In the cases we can use it for.

If you want automatic translation for a tourist for example it can get kinda wonky when referring to an object in front of you and the translation program doesn't know for example what gender the object is.
If you translate a book OTOH it will have to describe everything with words so the input is there.

Re: AIs are trained on grammatical sentences

[Zero__Kelvin]

We are machines and we can "do understanding" so clearly your claim that machines can't "do understanding" is proved already to be false. People said that everything we are doing today with technology could never be done. Claiming something can never be done is a claim that is as easily made as it is foolish to assert.

Re: AIs are trained on grammatical sentences

[gweihir]

Oh? And where do you take that certainty? Because Science very much does not say "we are machines". Science says "we have no clue how this works". Physicalism is religion, not Science. It has no scientific basis.

Re:AIs are trained on grammatical sentences

[gweihir]

Indeed. Good point.

Re: AIs are trained on grammatical sentences

[Zero__Kelvin]

I know what a machine is, whereas you clearly have no idea.

Re: AIs are trained on grammatical sentences

[Anonymous Coward]

Says the useless shit stain that doesnt know what the word machines means.

Re:AIs are trained on grammatical sentences

[hey!]

Translation from one human language to another is a different problem than understanding the semantics of natural language.

Human babies don't learn to translate English into some other language; they convert English into understanding. Learning to be fluent in a foreign language isn't learning to translate that language into your native language, it's learning to understand that language without translation.

So the ability of machines to translate from one human language to another does not at all look like what humans do. Again it's "AI" in the sense of replacing a human worker, but it doesn't shed any light on how that human worker would do the job.

Re: AIs are trained on grammatical sentences

[Anonymous Coward]

Godel begs-to-differ that humans are machines. In fact he proves we are not.

Re:AIs are trained on grammatical sentences

[religionofpeas]

Translation from one human language to another is a different problem than understanding the semantics of natural language.

I would agree for a crude translation. However, as you want to approach 100% accuracy in translation, understanding of language becomes essential. Look at how Google translates jokes or song lyrics.

Re: AIs are trained on grammatical sentences

[Zero__Kelvin]

That is a ridiculous claim. Knowing the name of a famous mathematician doesn't make you smart.

Re: AIs are trained on grammatical sentences

[SqueakyMouse]

If you're referring to his incompleteness theorem, then the same limitations it imposed on machines, it also imposes on humans. In much the same way a machine can't mindlessly rattle off closed tableaus of the negations of propositions until everything we want proved is proved, nor can a person. A human mathematician doesn't normally take this approach and is guided by intuition but there is nothing in his theorem to say a machine can't follow an algorithm indistinguishable from intuition.

Of course, whether or not you want to call a human being a machine depends on the definitions you're using and these definitions vary from context to context and from dictionary to dictionary.

Re: AIs are trained on grammatical sentences

[Anonymous Coward]

Oxford Dictionary (the highest authority on the English language) does not define "machine" as the human body. You are wrong and you're an idiot for not admitting it.

I think we might have a problem here...

[fuzzyfuzzyfungus]

It's good that we understand how humans acquire natural language well enough that 'just make the computer do it that way' is a plan. Otherwise this might not actually work terribly well.

Luckily AI is used to this class of failure by now, so they'll probably be OK.

Re:I think we might have a problem here...

[gweihir]

"AI" has failed to deliver on grande promises for half a century now. Nobody of those deciding about money seems to notice, so yes, they will be fine. The failure will continue though for a long, long time and maybe forever.

What has delivered a lot of results is classical, dumb automation. Calling it "AI" is just a marketing lie that seems to work well though.

Re:I think we might have a problem here...

[religionofpeas]

It's good that we understand how humans acquire natural language well enough that 'just make the computer do it that way' is a plan.

We don't understand how humans acquire knowledge of Go, yet people made a computer that started from nothing and learned it simply by playing itself and discovering all the knowledge.

The same method has been used in many different machine learning applications, and it seems to work pretty well, regularly scoring much better results than a human.

Re: I think we might have a problem here...

[Zero__Kelvin]

Finally someone who isn't a clueless baffoon pontificating on the impossibility of success in a field in which they have no education or understanding!

Re: I think we might have a problem here...

[dinfinity]

Indeed. Slashdot has become Luddite central, especially with regard to AI. The endless stream of empty "AI isn't actually AI" comments are boring and tiring.

Re:I think we might have a problem here...

[110010001000]

That is baloney. The Go/Chess/whatever playing computers didn't "start from nothing". They were PROGRAMMED to understand the game. The rest is marketing BS. They didn't just put the computer down and show it a bunch of Go games and it suddenly understood what Go was. Ridiculous. Computers are good at games with strict rules. We all know that. That is what computers are BEST at. You guys are just easily impressed and don't really understand technology.

Re: I think we might have a problem here...

[110010001000]

We aren't Luddites. We are people that actually understand technology and aren't just in IT.

Re:I think we might have a problem here...

[religionofpeas]

Computers are good at games with strict rules. We all know that.

The rules don't describe what a winning Go position looks like. And yet, that's what they learn to figure out, even better than humans.

Re: I think we might have a problem here...

[dinfinity]

Oh stop.

This is your utterly vapid comment, which is currently rated +5 Insightful:

MIT has been claiming this type of BS for decades. They haven't done anything. Literally they have been talking about this since the 1970s. Think about it: if it worked it would have been incorporated into something like Siri and be worth billions. But Siri is pathetic.

It adds nothing to the discussion but naysaying and bashing. It is a disgrace for the level of discourse I expect from Slashdot.

I am 'in IT' too, and I have been surprised by the speed at which self-driving car advancements have been made (I wasn't expecting us to be anywhere near where we are for at least 10 years -- go ahead, tell me you were). I've also been surprised at the effectiveness of deep learning in a project of one of our interns, whom I profusely warned about how ML can sound great but can lead to disappointing results (which it very much didn't).

That we are not at AGI level yet is obvious. The logic of the concept of AGI is that if we were, AGI would become ASI very, very quickly. I'm not going to predict when AGI is going to happen, but for everybody who actually understands technology it is clear that it will happen. All the "This isn't AI", "We don't even know how the human brain works" and "Siri is pathetic" comments are mostly just attempts at preservation of self-importance.

Whatever you want to call them, AI/ML/Expert Systems, those systems are absolutely slaying it and none of your Luddism is going to stop them.

Careful what you wish for

[Tablizer]

By breaking shit and seeing what mom and dad say about it?

Re:What a stupid question

[gweihir]

AGI, also known as "strong AI", "true AI", or "the AI we do not have and have not clue how to make or whether that is even possible".

Old flash game

[Anonymous Coward]

Many years ago, when flash was still hot, I played a game where you were on an alien planet trying to fix your ship. To accomplish this, you had to learn the alien language by observing and experimenting.

Does anyone recall this and know if it's still available?

results

[Ryan adiputra]

wow can it really be like that, then how will the results be seen

30 yrs away, just like fusion and self-driving car

[Anonymous Coward]

30 yrs away, just like fusion and self-driving cars.

"Expert" humans have a habit of underestimating the difficulties in solving some problems.

Having a huge lookup table isn't the same as real "learning."

It is just like how software developers claim some feature can be done in a week to 95%, but it takes another 6 months to get that last 5%.

In theory - yes, why not?

[dschiptsov]

Brain has evolved (pre-trained, if you will) machinery (particular structures of neurons, encoded in DNA) which has to be trained (literally) with actual human language. Think of an evolved deep-network which is trained to learn and maintain a deep-network for a language or two. The problem is - no one knows how to encode/represent this yet. Second order deep networks are unexplored.

Re:In theory - yes, why not?

[110010001000]

The human brain works nothing like a computer "neural network". The very fact that people call them "neural networks" is fraudulent.

Ignore Our Senses For Now, For Zorkâ(TM)s Sak

[turp182]

Stop with it with videos. Stick with ASCII text.

Machine learning, or AI, or whatever can learn patterns for things such as vision or hearing. But itâ(TM)s low level stuff like, âoeHey, look, a face!â. A newborn does this with baked in brain contents. âoeMake a low light picture look nice!â, a babyâ(TM)s eyes do this way beyond our photography tech (including AI assisted stuff).

Tangent with a point: Young children are sponges, they absorb everything they see, hear, smell, taste, and feel and note it in some fashion. They coalesce sensations into an understanding of the world. Attempting to simulate human sense response and understanding is why beyond our technical capabilities, excepting very specific scenarios that we target.

Photography improvements (commonly "AI") are simply pattern matching/task routines. They may be more complex than we explain, but they are task oriented with a super specific understanding area.

We should be focusing on AI research based around text and no other inputs. Words, semantics, descriptions of things. Describe see a sunset. Describe a scent. Describe having a feather brushed against an arm.

Describe the world. Donâ(TM)t try and show it to the technology.

Then let it try and play Zork.

I havenâ(TM)t finished reading this (searched for "ai to play zork"), but people are trying to solve parsed interactive fiction in this manner:
http://www.spagmag.org/issue-6...

The article also reminded of Minsky's Society of Mind. I've read that three times but never finished it.

Re:Ignore Our Senses For Now, For Zorkâ(TM)s

[turp182]

And don't use Microsoft Word to pre-type things for Slashdot. Slashdot's AI can't handle complicated text symbols like facing quote marks. Sad.

Except...

[Anonymous Coward]

...this isn't the way a child learns a language. Not really.

Children have teachers, someone who guides the child, corrects them, encourages them. Even when they aren't called teachers, that is the role they play (parents, friends, etc.). The learning process would undoubtedly be much longer, less accurate, and full of knowledge gaps without a teacher present.

The AI hounds always have this unstated goal of, "all we have to do is Load This File", or "all we have to do is Run These Simulations", or "all we have to do is Play These Videos", and Boom! The AI has the language (or other) skills we want. It all implies quick training runs, standardized inputs, reliable outputs, very factory oriented.

Except when we train children, we have to send them to school for 12 years, and that's a generally accepted practical minimum in a developed nation. Now imagine we had to train AIs the same way, sending them to expensive AI school for 12 years before they become useful.

Does AI technology seem set to take over the world in this scenario? Are we all imminently doomed to permanent unemployment? Could we be physically attacked or displaced or sidelined by a hostile or uncaring AI? Or any of the other popular Doomsday scenarios?

Or does it seem more like we have re-invented children, in the most expensive, least fun, and least rewarding way possible?

It's just nouns, verbs, some physics laws

[yes-but-no]

What is spoken language? it's just describing the visual objects and how they move about. Once you have object recognition, you need a set of verbs (standalone objects can shrink/expand; 2 or more objects: coming close together, separating away) -- these verbs define what the laws of physics/motion allow. Once you abstract these motions and got your verb set, describing the visual objects and their motion becomes what is spoken language.