naturalgradient

This makes sense to me as a characterization, thank you, although I don't understand why people get so upset about these distinctions as if I had personally offended them. Where I am from dragging the US into wars is naturally seen critically across parties, so far left is just about the issues you mentioned.

I did not call the NYT a far left magazine.

I said their reporting on EUROPEAN politics ("politics here") copies points from the far left (e.g. all mass migration is unquestionably good, parties against it must be right wing populists if not racists, etc).

I think their reporting on campus politics, identity politics is also far left, but other than that their stance on Iraq war etc is more Hillary-left than traditional 'left'. It's pointless semantics though, outrage mode is already engaged in this thread and it will probably soon turn into a dumpster fire.

Someone cannot just be wrong or inaccurate, they must be the enemy ('right rant'), and culture wars demand I first clarify I am on 'the right side of the issues' before saying anything. The more objective people think they are, the blinder towards their own bias. Of course I am biased too, but what people engage with in my post is the 'far left' comment on European politics instead of the actual point.

My point was that the NYT engages in culture war because it sells. I can agree with many issues on the NYT but still observe that and be annoyed by it, but that does not matter in tribalistic discourse.

To be honest I think you are just proving my point by getting upset and calling me 'ranty right'. Tribalism sells because it triggers emotions like these. It works for the Guardian just as well as it works for the NYT, Fox news or anyone else.

A good read on this is (at least to me as an outsider to American politics):

https://www.vox.com/2016/4/21/11451378/smug-american-liberal...

The greatest marketing trick the NYT has ever pulled off is presenting themselves as the last bastion of objectivity in the trump era.

Nothing could be further from the truth. I think the most recent Sarah Jeong controversy and virtually all reporting on migration, feminism, campus politics etc shows this. Mind you this is from a European perspective where I see almost all reporting about politics here as copying off talking points from the far left.

This is the true genius of their marketing though: They are actually as polarized as any other source in the culture war, but market themselves to an audience that likes to think of themselves as rational, objective, sensible.

The Washington Post. The NYT is unapologetically taking a side in the culture war and on many issues (campus politics, feminism, title IX, migration..) will ever only let one side make their point. The Wapo in my observation has a much better mix of both sides.

Your argument buys into the core of identity politics, namely creating a hierarchy of oppressors and victims, where the supposed victims can do no wrong. I find this an incredibly patronising worldview.

I would disagree with 3 because there are many in 2 and 3 which can go to 1 just by losing their job and e.g. having one high medical bill (in the states at least).

So surely there must be differentiation on whether any given job or emergency can affect your standard of living significantly.

You are referring to my comment about 'worth their salt' in a gross misreading. I said anyone worth their salt has other choices, so going to FB is a deliberate choice on these professors.

Any researcher in this area worth their salt can easily get cloud credit grants and collaborations from Google, Microsoft, Amazon. I pose if you go to Facebook, it's very much about money.

I understand the call of money but I cannot help but feel very negatively towards academics doing this with facebook of all organisations. After recent events, they cannot pretend not to know the impact and damage their work may have here. Excusing yourself with "I am just a researcher, I don't have anything to do with how my work is used" is just not good enough any more.

I would categorically reject any collaborations with FB as an academic in ML.

I completely disagree. You can find Google's pricing tactic aggressive, but it made me think about the entitlement there.

They complain that Maps at new prices would be more than the cost of their infrastructure, when actually their entire startup revolves around this data, and 5k to be able to use an amazing piece of high tech software that is ahead of the competition is..peanuts.

No matter what other business interests and strategies Google follows, there is no right to using such valuable tech for less than the monthly salary of an engineer. I find this incredible entitled.

The problem is that antitrust laws like these are very selectively enforced. Selection and prosecution of these cases is hence inherently political as these are non-standardized arguments or verdicts.

This is a weirdly shallow article containing lots of diagrams and bullet points to just summarize the known points that RL needs a lot of data and needs to learn from scratch.

No mention of all the ongoing work in learning from demonstrations, or more generally incorporating any off-policy knowledge. Vague speculations about the philosophy of model free learning. Not really worth the read (as someone working in RL).

Thank you this is a helpful data point.

Thank you for the response. It seems like liquidation preferences really come into play at growth stage when things get 'messier' due to capital needs, and declining them at seed round would send a very negative signal because the valuation must grow for any kind of success beyond the seed round?

Can someone familiar with the current funding climate say if standard deals at all levels involve liquidation preference nowadays? As in, if Im considering a seed-round, will there be any sophisticated investors doing no preference? Have talked to some investors in the scene (UK) but cannot seem to get a clear picture on this.

Is declining to accept a liquidation preference at seed level a red flag for any serious investor? What about subsequent rounds?

The guy said 'allow me to disagree slightly' after complimenting her on her thread and she lashed out at him in a highly unprofessional manner.

Thinking you can interact with customers of your employer like this in public even on a private account is naive.

Trying to turn it into an oppressor/sexism shitshow is being a toxic employee and a liability.

Hard to see a scandal..

There are schools in various countries, e.g. Romania, which are known for producing extremely well prepared applicants. They select for these schools from the entire country and have them focus on sciences, math and computer science early on. They practice Oxbridge style interviewing/math olympiad questions to death.

This results in for instance there being proportionally many more Romanians at Oxbridge Cs/Math/Physics than you'd expect by population.

Each applicant receives multiple interviews and I can perform one of these as a PhD student, and have received the same training any faculty member would have.

The interview process is relatively standardized, and if my results were to differ starkly from what more experienced interviewers do they would be disregarded, and the director of studies for that college would simply not invite me back to help.

I would add that asking PhD students to do this is not the worst thing because via supervisions and other teaching efforts, we have a good picture of what undergrads here need to be able to do. An interview is a like a short supervision.

I am aware, PhD students design and mark these questions, and can interview :)

The point is that if you go to a target school doing Math olympiads throughout your school life, the admissions exam and interview is a walk in the park. The applicants who didn't have any of this preparation can still do well but will fare relatively worse against that group, and I think this is very obvious in the ultimate intake.

>They try hard to let in disadvantaged kids, but it's a crap shoot deciding who has potential and who doesn't. They rarely care about anything other than academics. (I'm not sure how they recruit the rowers.)

They don't try that hard.

Being vaguely involved with the Oxbridge undergrad admissions process in CS/Math i can tell you there is very little trying. A fuss is being made about coming from a disadvantaged background but in practice sadly the people running it only care about one thing: how well you can grind out an answer to a math olympiad style question in 15 minutes. Yes, extra-curriculars and well-roundedness don't matter which I think is a good thing because I believe in focusing on being great at one thing.

What it comes down to nonetheless is preparation and school support, e.g. via training for math competitions. Saying the interviews are about 'evaluating the thinking process' of the applicant is a fantasy when most applicants come from schools where they have been trained to do them for years. Oxbridge are not forthcoming about this but ultimately they take people who are already well groomed Math olympiad winners, not raw potential.

It's probably still better than opaquely selecting for race and like-ability and if this means many math undergrads are Asian, why should that be a problem? It's still unfair to disadvantaged children and this sucks, but at least the criteria are clear.

Ps: on your question how they recruit rowers: They let them study land economy, that's the joke at least.

Thank you, I had not considered that. I think researchers or a HN audience may perceive this differently than the average reader.

So this goes off on a tangent but I feel it relates to noncentrality [0]. Fokas has a PhD in maths. Being an MD or having gotten an MD 40 years ago is clearly entirely non-central to his career. Calling him Mathematician-MD seems like it is meant to make him seem a lesser mathematician, e.g. by insinuating that this is just something he does part time, and that he can hence be taken less seriously.

I don't know what the poster meant by suggesting 'Mathematician-MD', but it reads weirdly to me for that reason. It's highlighting an attribute of a person that is entirely unrelated to his career or this article. Why if not to denigrate him? The title should be changed to neutrally reflect his position.

https://www.lesswrong.com/posts/yCWPkLi8wJvewPbEp/the-noncen...

Just pointing out that interestingly Cambridge, where Fokas is a professor, has not released anything.

He is merely visiting USC so it strikes me as weird that they would claim this PR so quickly.

Also Mathematician-MD somehow makes it sound like the MD means he is a lesser mathematician or not a full mathematician. Fokas is a well respected Professor at one of the top applied Maths departments in the world. A better and less biased title would be 'Math Professor' or 'Cambridge math professor' claims..

I think you should refrain from making such comments and follow the best-intention assumption on HN.

Yes, thank you for that by the way, did not want to diminish your efforts. Just wanted to point out that papers are often misleading about how many resources are needed to get to the point of running the result. I have received significant amounts of money from Google, full disclosure.

I would just want to comment that while this is true in principle, it's also slightly misleading because it does not include how much tuning and testing is necessary until one gets to this result.

Determining the scale needed, fiddling with the state/action/reward model, massively parallel hyper-parameter tuning.

I may be overestimating but I would reckon with hyper-parameter tuning and all that was easily in the 7-8 figure range for retail cost.

This is slightly frustrating in an academic environment when people tout results for just a few days of training (even with much smaller resources, say 16 gpus and 512 CPUs) when the cost of getting there is just not practical, especially for timing reasons. E.g. if an experiment runs 5 days, it doesn't matter that it doesnt use large scale resources, because realistically you need 100s of runs to evaluate a new technique and get it to the point of publishing the result, so you can only do that on a reasonable time scale if you actually have at least 10x the resources needed to run it.

Sorry, slightly off topic, but it's becoming a more and more salient point from the point of academic RL users.

Yes, I am aware, I did not mean random search as in random actions, but random search with improved heuristics to find a policy.

The point being that that the bells and whistles of PPO and other relatively complaticated algorithms (e.g. Q-PROP), namely the specific clipped objective, subsampling, and a (in my experience) very difficult to tune baseline using the same objective, do not significantly improve over gradient descent.

And I think Ben Recht's arguments [0] expands on that a bit in terms of what we are actually doing with policy gradient (not using a likelihood ratio model like in PPO) but still conceptually similar enough for the argument to hold.

So I think it comes down to two questions: How much do 'modern' policy gradient models improve on REINFORCE, and how much better is REINFORCE really than random search? The answer thus far seemed to be: not that much better, and I am trying to get a sense of if this was a wrong intuition.

[0] http://www.argmin.net/2018/02/20/reinforce/

Thank you for taking the time to respond, I appreciate it.

Well I guess my question regarding the expensiveness comes down to wondering about the sample efficiency, i.e. are there not many games that share large similar state trajectories that can be re-used? Are you using any off-policy corrections, e.g. IMPALA style?

Or is that just a source off noise that is too difficult to deal with and/or the state space is so large and diverse that that many samples are really needed? Maybe my intuition is just way off, it just feels like a very very large sample size.

Reminds me slightly of the first version of the non-hierarchical TensorFlow device placement work which needed a fair bit of samples, and a large sample efficiency improvement in the subsequent hierarchical placer. So I recognise there is large value in knowing the limits of a non-hierarchical model now and subsequent models should rapidly improve sample efficiency by doing similar task decomposition?