AI Trained To Navigate Develops Brain-Like Location Tracking (arstechnica.com)

An anonymous reader quotes a report from Ars Technica: Now that DeepMind has solved Go, the company is applying DeepMind to navigation. Navigation relies on knowing where you are in space relative to your surroundings and continually updating that knowledge as you move. DeepMind scientists trained neural networks to navigate like this in a square arena, mimicking the paths that foraging rats took as they explored the space. The networks got information about the rat's speed, head direction, distance from the walls, and other details. To researchers' surprise, the networks that learned to successfully navigate this space had developed a layer akin to grid cells. This was surprising because it is the exact same system that mammalian brains use to navigate. More DeepMind experiments showed that only the neural networks that developed layers that "resembled grid cells, exhibiting significant hexagonal periodicity (gridness)," could navigate more complicated environments than the initial square arena, like setups with multiple rooms. And only these networks could adjust their routes based on changes in the environment, recognizing and using shortcuts to get to preassigned goals after previously closed doors were opened to them. The study has been reported in the journal Science.

Comments

"solved" my ass

[Khashishi]

To solve a game means you know the optimum move to make (and therefore know the winner of the game from the start). Deepmind has only defeated humans, not solved the game.

Re:"solved" my ass

[mikael]

What happens when it plays a game against itself? That's always the fun thing to do when playing with AI.

Re:"solved" my ass

[Kjella]

What happens when it plays a game against itself? That's always the fun thing to do when playing with AI.

That's all AlphaGo Zero did for training, play Go against itself. They released some of the games in the final configuration, they're extremely hard for humans to understand, like a novice chess player who doesn't understand how the grandmaster moves affect the game ten moves down the road. Watch some of the live commentary/broadcast on Lee Sedol's second game, move 37... they're SO confused, some wondering if AlphaGo has gone off the rails. But as the game progresses it becomes very clear it's seen further than all of them and is the setup for a crushing attack. And that was a much earlier iteration that would be completely crushed by the latest one. They'll certainly be studied but it's uncertain how it's highly uncertain how much those AI vs AI mind games tell anyone else.

Re:Just like the human mind!

[fluffernutter]

I was thinking "Roomba" myself: "How does that thing find the little dock??"

Re:Just like the human mind!

[110010001000]

Interesting. "Roomba" powered by DeepMind.

Obviously all male team.

[nospam007]

Half the Natural Intelligence on the planet knows that if you just ask for directions will get you faster there than playing with gizmos.

Evolution

[duke_cheetah2003]

Pretty fascinating that evolution takes the same path, biological or mechanical, when presented with the same challenge situation. Super fascinating. Makes you wonder about mammals, if a computer is evolving the same mechanisms as real-world life did.

Re:Evolution

[Dagger2]

You say that as if you think that our brains aren't just weighted directed acyclic graphs -- but that is in opposition to the article and the research, which seems to demonstrate that that's pretty much what brains are, at least for this particular part of their functionality.

Re:Evolution

[raftpeople]

Disagree. This is an interesting result, that the physical structure of the network (as seen in rats and used in this neural net), outperforms the network that is not structurally similar to the thing that it is modeling (e.g. 3d or 2d space).

Neural nets are universal function approximators, but that doesn't mean that all nets are equal. Mapping the structure of the net to the problem being solved to achieve better results is shown to be valid and important, and nature obviously discovered this as well.

Re:Just like the human mind!

[fluffernutter]

The l'il vaccuum that returns to it's dock and proceeds to kick ass at Go on the internet.

Blame the "journalist"

[HeckRuler]

Yep, that's a pretty shitty scientific-journalism article. Go isn't solved, AI players now beat human players. Throwing out a definition of "navigation" like anyone doesn't know what it is. Claiming we know how brains navigate. And I really almost missed the link to the ACTUAL paper at the bottom.

But the paper? The paper doesn't jump out as immediate shit. For example, apparently we DO know how mammalian brains navigate:

[AI fails] to rival the proficiency of mammalian spatial behaviour, which is underpinned by grid cells in the entorhinal cortex (Hafting, T., Fyhn, M., Molden, S., Moser, M.-B. & Moser, E. I. Microstructure of a spatial map in the entorhinal cortex. Nature 436, 801–806 (2005).)

huh.

And then these people tried to mimic that. "Here we set out to leverage the computational functions of grid cells to develop a deep reinforcement learning agent with mammal-like navigational abilities". It's not that AI developed this, the researchers specifically developed AI to have it. The research is real. It's kinda interesting. It's nothing ground-breaking (and actually seems kinda.... college-student level... but I dunno) but is another incremental improvement in the field of study.

The fuckwit journalist who butchered it (that's DIANA GITIG - 5/11/2018, 5:15 AM btw) needs to stop with the hype bullshit. It diminishes arstechnica as a trash rag that can't even report on pop science.

Re:Blame the "journalist"

[RespekMyAthorati]

It diminishes arstechnica as a trash rag that can't even report on pop science.

How can anybody diminish arstechnica any more than it already is?

0% journalism, 100% clickbait.