Starred posts

nbro

May 24, 2021 15:46

Questions of the form "How do I implement this thing in TensorFlow?" or "Why am I getting this IndexError?" are off-topic. A question of the form "Why am I getting this training and test curves? Why isn't the test curve always decreasing", or something like that, and the OP also includes code in the post, I think that would be fine, because the problem is probably conceptual (i.e. overfitting)

DukeZhou

May 10, 2021 01:33

2

A: Should the threshold for post closure by non-moderators be reduced from 5 to 3?

Some of you may have noticed the Meta Stack Exchange post - Testing three-vote close and reopen on 13 network sites (it's linked in the featured on meta sidebar) - we've finally got this project under way and Artificial Intelligence is one of the sites we'll be running the test on. Starting tomor...

DukeZhou

Apr 29, 2021 02:05

I think we mods should encourage participants by at least upvoting valid, in-scope question, and, where possible, upvoting good answers. (The latter requires more time to carefully review the answer, but the former requires only parsing.)

Neil Slater

Apr 18, 2021 08:39

If you use more specialist architectures such as CNNs, then that could be 100s of 1000s inputs.

Neil Slater

Apr 18, 2021 08:38

It is not uncommon to train MLPs with 1000s of input neurons.

Neil Slater

Apr 18, 2021 08:35

There is no theoretical limit on the number of inputs p. The limits will be practical:

DukeZhou

Oct 21, 2018 22:23

@nbro nice link. thanks for posting:

PREPARING FOR THE FUTURE OF ARTIFICIAL INTELLIGENCE
Executive Office of the President National Science and Technology Council Committee on Technology, October 2016
https://obamawhitehouse.archives.gov/sites/default/files/whitehouse_files/microsites/ostp/NSTC/preparing_for_the_future_of_ai.pdf

2

Nike Dattani

Feb 15, 2021 08:30

Exactly 38000 users!

AI Meta Questions

Sep 26, 2016 17:22

1

Q: Moderator Pro Tem Announcement

Throughout the beta, we need members from the site whose focus is to engage the community, both in community-building issues and site management. That's why we select a few members from each community to act as temporary, provisional moderators. You can read about the program here: Moderators Pro...

3

RewCie

Feb 6, 2021 04:12

.

nbro

Jan 25, 2021 20:33

Ben N

Aug 16, 2016 15:21

@AIMetaQuestions A moderator nomination post is very good news! It bodes well for our chances of continuing into public beta

3

AI Meta Questions

Aug 16, 2016 13:32

1

Q: Accepting Nominations — Who should moderate this site?

Ideally Moderators are elected by the community, but until the community is large enough to hold a proper election, we will be appointing three provisional Moderators to fill those roles. We need your help. Please nominate folks you would like to see become provisional moderators for this site. ...

discussion featured moderation

3

RewCie

Dec 31, 2020 13:35

Happy New Year everyone! Year of Disruption in AI begins here :-)

DukeZhou

Oct 28, 2020 01:08

"In a sense, most functions are uncomputable: there are Aleph-null computable functions, but there are an uncountable number (2^Aleph-null) of possible Super-Turing functions."

[J. Cabessa; H.T. Siegelmann (Apr 2012). "The Computational Power of Interactive Recurrent Neural Networks"]

DukeZhou

Mar 9, 2018 18:27

Thanks to everyone whose been pitching in with editing tags, questions and answers. Apologies if it sometimes takes a while to approve--have to check everything. But it is noted and appreciated, and I'm glad to to see the community spirit in striving making this Stack as useful as possible!

2

nbro

Jul 10, 2020 22:12

Why do deep convolutional networks generalize so poorly to small image transformations?

AI Meta Questions

Sep 22, 2017 16:32

1

Q: Please welcome another pro tem mod!

I'm Pops, a Community Manager at Stack Exchange. Though it once again saddens me to say it, one of your moderators has decided it's time to step down. Another of your fellow community members here on AI SE has answered the call to be your new pro tem mod, though: Please join me in thanking Mat...

discussion featured moderators

2

DukeZhou

Jul 30, 2017 21:19

I'm reaching out partly because you felt strongly enough to critique the scope of this Stack, and your censure stuck with me--now that I'm a pro tem mod I may be in a position to effect change. I have also found your questions on Math to quite useful, which, from my perspective, adds weight to any ideas you might have on how AI and the hard science stacks need to interconnect.

2

Feeds

May 5, 2020 14:18

The Berkeley Artificial Intelligence Research Blog: Four Novel Approaches to Manipulating Fabric using Model-Free and Model-Based Deep Learning in Si…

posted on May 05, 2020

Humans manipulate 2D deformable structures such as fabric on a daily basis, from putting on clothes to making beds. Can robots learn to perform similar tasks? Successful approaches can advance applications such as dressing assistance for senior care, folding of laundry, fabric upholstery, bed-making, manufacturing, and other tasks. Fabric manipulation is challenging, however, b…

DukeZhou

Dec 9, 2019 18:32

In this case, the user in question has received an initial 7-day suspension. Stack believes in second-chances, so it is our hope that the user will self-moderate and restrain their activity in future, to avoid further action.

Neil Slater

Nov 29, 2019 19:14

"Machine learning is the science of credit assignment." Seems like a familiar writing style :-)

nbro

Nov 29, 2019 17:01

According to Jürgen Schmidhuber, Le Cun, Bengio & Hinton (LBH) fail to credit the pioneers of the deep learning field, which originated half a century ago, while they keep on citing each other and call themselves "deep learning conspiracy", which is completely ridicolous. Alexey Grigorevich Ivakhnenko is the actual father of deep learning.

John Doucette

Nov 19, 2019 02:10

@nbro Done. Thanks for finding that, and for raising all those close votes. There was only one I disagreed with, which I think could be a good example of a "programming" question that I think is on-topic (ai.stackexchange.com/questions/16557/…). Although the poster includes code, and the code is using a modern library, the problem the asker is having is really rooted in an AI-related misunderstanding.

nbro

Nov 5, 2019 16:41

Here's a very good question, IMHO.

DuttaA

Nov 1, 2019 01:11

technologyreview.com/s/614597/…

DukeZhou

Oct 22, 2019 19:14

I'm asking people to take the time to read and comment so we can get a sense of what the community would like to see.

DukeZhou

Oct 22, 2019 19:11

Essentially, we've come a long way in the past few years, but it may be time to start getting leaner and stricter to take the Stack to the next level.

DukeZhou

Oct 22, 2019 19:10

SE:AI users, please take the time to review and comment on this well considered answer. (I agree with almost everything):

Eric Platon

Aug 12, 2016 01:25

Call for participation:
Less than 2.5 answers (I like half-answers) is not good for the private beta. I am trying to post second answers here and there, sometimes just to add some perspective or extra information (which I think is quite positive).

Could you do the same on topics of interest? One time-consuming aspect is to write at least fair quality contents, and if you choose to participate on this call, please keep it in mind.

2

wythagoras

Aug 11, 2016 12:17

Did some of you already see this one: xkcd.com/1613

2

Dawny33

Aug 6, 2016 14:50

@BenN We do have exclusive content (<-- The posts you linked). Just don't close off posts which ask about algorithms of AI systems. If we keep closing them, we'd be closing posts all day long, my friend.

And it would definitely discourage users (like me) who want to understand the fine details about AI systems :)

2

Ben N

Aug 6, 2016 14:46

Right now, the default state is fail unless we can show SE that we bring something new to their network

2

user76284

Sep 1, 2019 18:12

@nbro Done.

DukeZhou

Aug 9, 2019 19:06

DukeZhou

Aug 9, 2019 18:55

Three pitfalls to avoid in machine learning - Nature

DukeZhou

Jul 30, 2019 18:05

@nbro good point. "relation" has a specific set theory meaning, so potentially confusing to use informally. Relation has been merged into comparison, and relation deleted.

Dennis Soemers

Jul 8, 2019 07:13

Debugging / checking correctness of RL implementations, and reward design, are both notoriously difficult. RL can often still appear to work "fine" (but maybe not great) when there are in fact a few bugs. And sometimes the implementation will be perfect and it'll still not work well :D I think the same kinda holds for ML in general, not just RL.

Anyway, for debugging, there are a few good tips in these slides (I think the talk itself is also up on youtube): joschu.net/docs/nuts-and-bolts.pdf

Neil Slater

Jul 7, 2019 10:56

@DuttaA: I haven't done anything sophisticated, but assuming you are using a normal supervised learning algorithm internally that you trust (e.g. using Keras to train your value and policy networks), then you are mainly concerned that the algorithm can find a good approximation to optimal policy, and that any value-based components report expected value accurately

Neil Slater

Jun 17, 2019 06:31

@DuttaA: I am not 100% certain that I did recommend AWS directly. Nowadays, if you know how to use a Jupyter notebook to arrange your work, then for PyTorch I'd recommend Google's Colab, which is completely free, and easier to setup and use than AWS.

DuttaA

Jun 7, 2019 08:13

technologyreview.com/s/613630/…

DukeZhou

Jun 3, 2019 17:26

JD Hamkins, one of the authors of Transfinite Values is a regular contributor on Math Overflow!

DukeZhou

Jun 3, 2019 17:15

@DuttaA The problem with Chess is it's purely positional, and not a simple path-findey way like with m,n,k-games. You can't get a precise estimation of value, even of a given game piece, so the structure of the game makes simple heuristics problematic. (Deep Blue was a massive undertaking, and both it and Stockfish rely heavily on game books--libraries of expert games.)

Mez13

Jun 3, 2019 13:05

@nbro Thank you!

Mithrandir

Jun 2, 2019 18:01

Chat rooms are often a place where those who are most invested in the site will congregate, and they're often the ones who are most invested in the moderation of the site - such as voting, closing, and reopening. So it's really not unusual to see discussions about such topics in the chat rooms.

DukeZhou

May 13, 2019 18:43

The Lottery Ticket Hypothesis: Finding Sparse, Trainable Neural Networks

Manuel Rodriguez

May 7, 2019 22:43

youtube.com/watch?v=F3NhxxG9F68 Gordon Ramsay Tearing Into Amateur Chefs | Hell's Kitchen

Philip Raeisghasem

Apr 28, 2019 18:31

@ManuelRodriguez Not trying to gang up on you here, but I think @nbro has a point. When your answers don't make sense to other established users (or at least seem to use unorthodox terms for things), it's not unreasonable to be asked for sources. Peer review is impossible if we don't speak the same language or share common references.

Jan

Mar 20, 2019 21:23

Hi All, I am trying to get started with reinforcement learning where I only get a reward after playing an episode. Does anyone know some links or resources for this?

DukeZhou

Mar 18, 2019 21:40

There is a certain poetry to "Since AIXI is incomputable, it assigns zero probability to its own existence."

The Singularity

The Singularity