[rwnspace]

I sense the hand of an editor. Particularly regarding the title.

Embodiment seems to be a branch with low-hanging fruit, when it comes to advancing AGI. I think the economic structural problems are important, but it's possible to over-egg the details and for some lab to stumble on an experimental paradigm with features we didn't realise were implicated a priori. When it comes to other AIs, the idea that we are stuck for pragmatic/practical issues is a little silly.

I'm no expert, just a person with an arm-chair (and too much time on my hands), but I suspect that idealising the feature-space we work with can hide as many things as it reveals - it may turn out that the computational problems are so large because we are mostly attempting to solve them ex nihilo. That is, embedding in an environment plays as much a role in the process of intelligence as a neuronal structure does; genes and evolution provide a mode for translating environmental computation into neuronal computation. The vast scope of what we don't know about the role of glial cells for cognition (and the little that we do) makes me doubt that complex structures of binary mechanisms will be sufficient. But again, that's just my speculation, and perhaps lack of education.

[gwern]

> I sense the hand of an editor. Particularly regarding the title.

Gary Marcus is probably fine with the title. He's been talking down deep learning (and talking up his own more old-fashioned Bayesian flavored ideas and startup) for years now, trying to ignore all the successes like Google's knowledge graph and just omitting the actual research, like when he says

> Such systems can neither comprehend what is going on in complex visual scenes (“Who is chasing whom and why?”) nor follow simple instructions (“Read this story and summarize what it means”).

Which deep learning actually... works pretty well on? Look at Facebook and Google's work on visual & textual question answering using approaches like memory networks.

[alexcnwy]

Totally agree

The work on textual understanding is definitely still early days (though the press release makes it seem like it's understanding the entire LOTR, it's just reading a very structured short version) but already crazy impressive: https://venturebeat.com/2015/03/26/facebooks-latest-deep-lea...

On visual QA, Francois Chollet's talk at the TensorFlow dev summit shows how easy it is to get a (again constrained to 1-word answers but still very impressive) video QA system working in like 20 lines of Keras: https://www.youtube.com/watch?v=UeheTiBJ0Io&vl=en

[ethbro]

The person you're largely agreeing with is Rodney Brooks (formerly of MIT CSAIL, http://people.csail.mit.edu/brooks/publications.html ).

And to sum it up a bit, the hypothesis is that humanlike AI is as much a product of the experience and reality of being physically (and limitedly!) human as it is any abstract algorithm.

You may also know him from a small company called iRobot (aka Roomba).

[rwnspace]

Thank you very much for the link, and nice summary. This looks very interesting, I'm grateful: my learning habits mean I'm liable to playing with second-hand scraps of ideas, and missing out on sources.

[arstin]

I think his article Intelligence Without Representation is particularly worth reading at least once.

http://people.csail.mit.edu/brooks/papers/representation.pdf

[ethbro]

To elaborate a bit, most of the excerpts I've heard have to do with the availability of sensors.

E.g. Task: grasp an egg without cracking it

Physical platform 1: actuators, no pressure sensors

Physical platform 2: actuators, pressure sensors where egg contacts robot

Inarguably, the simplest successful implementation of the task in driving code will be much more concise for platform 2 than platform 1.

... Now generalize the same idea to trying to teach disembodied AI to be human.

[visarga]

> Embodiment seems to be a branch with low hanging fruit, when it comes to advancing AGI.

If it were "low hanging" it would have been picked already. Reinforcement learning with AI agents is hard, especially in a dynamic environment with many types of objects.

I think the path towards AGI is to do simulation coupled with deep learning. Simulation would open the door to predicting non-trivial effects that cannot be learned by example because they are so rare that there are no training examples. We can generate artificial training examples to cover all the rare cases.

[rwnspace]

I'm sorry if 'low-hanging' comes off as disrespectful - I'm just guessing that aspects of embodiment, once understood, will be capable of fairly trivial description and reap large consecutive rewards. I remember your u/n from other posts, we seem to have similar interests but you are vastly more educated in the engineering of AI. What's your background, if you don't mind sharing here?

I am suspicious non-contingent aspects to cognition remain that simulation and deep learning don't necessarily grant, though they might well be sufficient. I'm not smart enough to be sure, and I'm stretching for a description: a child self-reared to adulthood in the wild won't display what we usually consider essential facets for 'humanlike' levels of intelligence or competence. We're hardly trying to build a caveman.

They lack whatever is crucial in socialisation -- the ability to make subtle differentiations between other agents' actions and motivations seems to endow self-awareness, and abstractions for successfully handling novel objects and ordering perception relevance. Successful generality to our degree seems to be better 'outsourced' rather than hard-coded into solo agents, at least in the natural examples. Though I understand that's not necessary, perhaps there are good reasons for it. I feel like the first AGI will actually look a lot more like "multiple similarly 'perspected' AIs interacting with one another leads to each carrying the G in AGI". Essentially I'm suggesting it's hard to have generality and relevance to our proficiency (or better) without a 'culture'.

What I'm thinking seems to boil down to inserting some of Piaget's ideas into the philosophy of AI, which might be a bit much, and I'm open to charges of bullshit.

[visarga]

I am just a hobbyist in machine learning. The current situation in AI is that we can only do limited aspects of perception: like vision and hearing, at a superficial level. We can recognize objects but we can't recognize relations between objects nearly as well. There is no global scene understanding yet. With text, we can do syntax and translation, but we can't do reasoning except on trivial cases. There is no power of abstraction yet, of transferring knowledge between domains, which is essential for advanced intelligence.

So, before we have an embodied agent, we need to solve the reasoning and abstraction part, and my money is on graph signal processing (a kind of neural nets) and simulators (also implemented as neural nets). We need to move from simple object recognition to reasoning and simulation on graphs of objects and relations.

[cscurmudgeon]

> We can generate artificial training examples to cover all the rare cases.

That means we have a model to start with. Chicken & egg.