Grid-Cell Like Deep Layers in Navigation AI | Hacker News Reader

Grid-Cell Like Deep Layers in Navigation AI

nature.com

77 pointsSubiculumCode28 comments

This DeepMind work has been gathering a lot of hype, but it is very similar to this work by Cueva and Wei:

Published on Arxiv on March 31: https://arxiv.org/abs/1803.07770

Published at ICLR 2018, a well-known machine learning conference: https://openreview.net/forum?id=B17JTOe0-

It is rather disheartening that they didn't acknowledge the -- seems like a deliberate oversight.

The Cueva and Wei paper is well written, explains concepts clearly, without the hype, and published in an open access venue.

Yes, this DeepMind paper has been associated with an extraordinary amount of hype! This must be the third or fourth time it has been posted here, when most of the extensive grid cell literature has never been posted once; including theoretical work that explains the emergence of such cells, or that uses grids for navigation. And I don't think this is the first time that a DeepMind paper has been selective in its literature review, either...

Was the arxiv paper ever peer reviewed?

Yes, refer to the openreview link to my earlier comment on this thread.

Well, the paper was received from Nature on 05 July 2017, so it isn't surprising that they didn't cite the work from Cueva and Wei published 21 Mar 2018.

If you are making that argument, you should refer to the ICLR deadline, which is well earlier, but still after July 2017.

That's not quite how the process works. These papers go through multiple (> 2) revisions. At any iteration, there's ample opportunity for updating references. This applies double given how long the Cueva/Wei work has been available (preprint & CCN'17 (?) contribution).

I've noticed that DeepMind completely ignores imitation learning area of research. I never saw them citing any of the fundamental papers, despite a bunch of their reinforcement learning ideas explicitly match those ideas that are now 8+ years old.

Additional context for the overlap with previous work on grid cells in neuroscience:

https://twitter.com/gershbrain/status/994678940043284480

Thanks for point out this!

I'd rather read the similar work from ICLR than nature, which I do not have access...

The link says I have to pay $8.99 to "rent" the article, or subscribe every year for $199. WTF? Is this science?

Why can't people publish research in open access journals or conferences?

Unfortunately a lot of researchers submit themselves to closed access journals due to prestige factors. I understand why a PhD student or academic might do this, to rake up impact factor points in order to graduate or get pass the tenure committee. But these guys are DeepMind, with high salary jobs already, without the academic BS, so you'd figure they would know better ...

This is Nature, the single most prestigious journal on science. It’s the Ivy League of PhDs. A publication in that journal is a guaranteed job safety in university circles.

maybe if you are a biologist/neuroscientist. but for machine learnign research, how sure are we?

We are. Nature has enormous prestige across all fields. It has the highest H-index.

Nature (the multidisciplinary journal) should not be confused with other Nature publications that have the name "Nature" in them, like Nature Neurscience or Nature Reviews Molecular Cell Biology.

What do you mean? H-index applies to individual researchers. All nature journals have high impact factors - and readership among hard scientists, but i dont think ML researchers read Nature every month. I understand that some reserachers need the prestige factor. But honestly , in the case of deep mind, i think they are giving prestige to Nature as an ML journal, where they previously had none. They could have published their results in a reputable open access journal like eLife instead, and do a service to the open access community. Publishing closed access is not a great direction to move towards. Also, why didn't they pay for Nature's open access option, it's not like they don't have the money.

> What do you mean? H-index applies to individual researchers.

H-index is also applied and calculated to publications (Journal h-index). IMHO It fits better to the multidisciplinary nature of the journals than other impact factors.

>some reserachers need the prestige factor. But

In academia things like Nobel price, Abel price, Turing price are signals of significance for those not in the field. They are usually received very late in career.

If you have several publications in Nature, the prestige level approaches those awards and you get the prestige when you are still active researcher. You get tenure or funding for research group. The value can be measured in millions.

i think nobel winners get more nature publications post-nobel than before

I have yet to see ML phds at my university (a well reputed ML institute ) care for Nature. Everyone publishes in journals specific to their area.

* Nips, icml, iclr for all ML folk * Cvpr, iccv for vision * Nacl for nlp * Siggraph for graphics

So on and so forth.

I wonder how nature finds reviewers for ML papers , given its generic nature. Even ML researchers themselves have been struggling to give good reviews in the aforementioned top conferences. I can't imagine how the reviewers at Nature could appropriately review papers of the same quality.

I would say the fact that machine learning / neural network theory is featured in Nature is quite a breakthrough. And neuroscience for that matter is quite an associated field of research.

Friend of mine recently published a paper in Nature on quantum computing. There is immense prestige in doing so.

no doubt, but the question is at what cost and for what reason. Nature editors are at best an arbiter of significance, a job that they often delegate to voluntary external referees anyway. Nature is not a speciaalized journal. Like Science, they offer prestige but bundling all the sciences in one journal does not make them useful for regular reading. It's a self-selecting process, and scientists who are open-science-minded should retarget this self-selection process to one of the good open access publishers. Why should it be Nature that gets the spotlight here and not eLife, an open publisher that is truly innovative.

Deepmind has an unfortunate history of publishing in Nature for the prestige. Maybe because Hassabis was a neuroscientist and acquired that bad habit. But does it matter in the CS/ML world? I think it doesn't, because a number of people have done it with papers that range very widely in significance (e.g. this is much less significant than the original deep q learning paper that was also in nature iirc)

From my observation (from knowing a few people there from my lab), DeepMind is still organised relatively similarly to university groups.

Professors moving over are made to lead research teams, and while there may be some overall organisation goals, researchers still care about their individual career advancement and prestige as measured by paper output. So individually, most people working there are incentivised towards doing what is best for their individual prestige in the research community for after leaving deepmind.

While it doesn't matter that much to machine learning folks, Nature papers are a huge flag to take the work seriously for journalists, science enthusiasts, and researchers outside of computer science. DeepMind's ICLR and NIPS papers receive much less media attention than their Nature papers.

This is not a huge problem anymore, because scihub exists. For academics it also doesn't matter because they have a university library subscription for all the prestigious journals.

It certainly matters because the whole Machine Learning field was pretty much 100% open access before DeepMind “went big” and started pursuing Nature publications from the Atari paper and onwards. I completely understand their move as Nature brings with it a lot of prestige well beyond your own field, it is a genius tactical move. But I still feel that this undermines something rare and precious that we have as a field.

A free "unformatted" version [1] is linked from the bottom of their blog post [2], as with their other Nature papers.

[1] https://deepmind.com/documents/201/Vector-based%20Navigation...

[2] https://deepmind.com/blog/grid-cells/

Hiearchichal Temporal Memory is exactly what this is. HTM has been successfully used to create artificial hypocampus. Watch the video in the first link below and try not to be impressed.

Links:

https://discourse.numenta.org/t/oscillatory-thousand-brains-...

https://en.m.wikipedia.org/wiki/Hierarchical_temporal_memory

Reply on news.ycombinator.com