czr

This title ('The infamous “AI gaydar” study was repeated – couln't be reproduced') is very incorrect; it does not correctly summarize either the linked article or the actual paper (https://arxiv.org/pdf/1902.10739.pdf) that the article is discussing.

From the abstract on the actual paper:

> Using a new dataset of 20,910 photographs from dating websites, the ability to predict sexual orientation is confirmed... While demonstrating that dating profile images carry rich information about sexual orientation these results leave open the question of how much is determined by facial morphology and how much by differences in grooming, presentation and lifestyle. The advent of new technology that is able to detect sexual orientation in this way may have serious implications for the privacy and safety of gay men and women.

So, the study was repeated, and could be reproduced. However, it's still not clear if there's any merit to the facial morphology hypothesis, vs. the (less worrying) hypothesis that the models are just picking up on differences in grooming and presentation.

From the original paper (http://cnnlocalization.csail.mit.edu/Zhou_Learning_Deep_Feat...):

"A class activation map for a particular category indicates the discriminative image regions used by the CNN to identify that category"

So, it's mostly what you said, but the maps you get are specific to the particular trained network you're using; the class activation map tells you, for a given network, for a given image, for a given class, which image regions did that network consider relevant to that class.

For learning more about interpreting neural nets I would highly recommend the posts on distill.pub; https://distill.pub/2018/building-blocks/ and https://distill.pub/2019/activation-atlas/ are probably most pertinent.

Further reading:

* Blog post: https://openai.com/blog/openai-five-finals/

* Live tweets: gdb at https://twitter.com/gdb/status/1117106896610058240, psyho at https://twitter.com/FakePsyho, mtrc at https://twitter.com/mtrc, nickstatt at https://twitter.com/nickstatt

Some tests:

Easy: https://i.redd.it/dwi7jrmrwri21.jpg -> https://i.imgur.com/bQZXrAJ.png (Okay, but doesn't mask shadows around the feet correctly)

Medium: https://500px.com/photo/195143459/ximena-by-agata-serge -> https://i.imgur.com/kUdcvFo.png (Okay; it could have taken out a bit more background in some semitransparent sections of hair)

Hard: https://i.redd.it/i7fn0fy89ma21.jpg -> https://i.imgur.com/WGM67XL.png (Fooled with some false-positives, but still a good try)

Certainly not perfect, but definitely usable.

>To the untrained eye, the code of a computer virus is just a jumbled mess of letters, numbers and symbols.

>But to Fabian Wosar, each line is a clear instruction. He knows and understands every digit and dot in the same way a pianist would read a page of musical notes.

This stood out to me as a strangely flowery description of reading asm, but apart from that the article was quite engrossing. Props to Fabian.

It indicates that you think the work is a meaningful contribution towards (but not a full solution for) the problem named by the paper title. I'm not sure if there's a distinct origin (see https://english.stackexchange.com/questions/188988/what-is-t...), but it's fairly conventional these days.

Certainly if they make a strong AGI, money is no longer an issue. I'm curious what they will do in the, ah, interim, on the off-chance that inventing SAGI turns out to be a difficult problem.

Is there any indication of what avenues OpenAI will be (or would consider) using to generate revenue? A lot of the most financially lucrative opportunities for AI (surveillance/tracking, military) are morally ambiguous at best.

Indeed. There's also a slightly less charitable interpretation–Tesla could't ship every car with LIDAR for cost reasons, so they're strongly incentivized to claim they don't need it.

That said, I don't think Tesla are wrong. If Tesla can deploy MobilEye-level automatic visual mapping (see https://youtu.be/GQ15HWCw_Ic?t=1381) this can obviate the need for LIDAR-based localization and dramatically improve their perception systems (by having very good prior information about all static obstacles, such as lane dividers).

Dynamic obstacle detection will still be worse than it would be with a vision+LIDAR approach, but not that much worse. Having superhuman ability along two axes (map-based priors and alertness / reaction time) is likely sufficient to drive better than humans, even if several remaining axes (dynamic obstacle detection and prediction) are worse.

Additionally, Tesla is not (AFAIK) using any structure-from-motion (https://www.youtube.com/watch?v=KT2KsN7yKo0) or stereo-vision (https://www.youtube.com/watch?v=SskSDjUG8ZY) techniques, but there's a chance that these could also improve perception, especially in tough cases where a single frame is not enough for good detection (e.g. white semi in sunlight).

LIDAR is not being used in the (compared to Tesla) small-scale AV systems of Waymo, Cruise, Aurora et. al because it's essential, but rather because it's convenient, and because the companies producing those systems want to give their AVs every advantage at any cost. I fully expect Tesla (and MobilEye) to do superhuman self driving without LIDAR, but it will take longer (i.e. not by EOY).

The context is that Tesla has had troubles in the past, and will continue to have troubles into the future, so this particular article is not indicative of any inflection point. It's the continuation of a long trend of mismanagement, shifting timelines, and questionable decisions that have nonetheless failed to take the company down.

For context,

"An Electric Carmaker Struggles as Its Production Lags" (https://dealbook.nytimes.com/2012/09/25/questions-about-elec...) - September 25, 2012

"Stalled Out on Tesla’s Electric Highway" (https://www.nytimes.com/2013/02/10/automobiles/stalled-on-th...) - Feb. 8, 2013

"Lurching Start for Tesla in China" (https://www.nytimes.com/2015/02/11/business/international/te...) - Feb. 10, 2015

"For Tesla Owner, Losing a Wheel Was Just the First Surprise" (https://www.nytimes.com/2016/06/11/business/tesla-motors-mod...) - June 10, 2016

"Tesla Model 3, Elon Musk’s Grail, Remains a Costly Pursuit" (https://www.nytimes.com/2018/02/07/business/tesla-musk.html) - Feb. 7, 2018

"Tesla Looked Like the Future. Now Some Ask if It Has One." (https://www.nytimes.com/2018/03/29/business/tesla-elon-musk....) - March 29, 2018

The headline here is not a good description of the actual paper (https://arxiv.org/pdf/1902.11097.pdf).

The difference in average precision in bounding boxes between the LS (light skin) and DS (dark skin) observed by the researchers was around 1-10%. Not "can't recognize". Notably, people with dark skin were underrepresented in the dataset that the researchers used.

When they compensated for the dataset imbalance by reweighting examples, the issue was resolved.

Notably, the research itself is on publicly-available object detection systems, not any actual autonomous vehicles. Any company seriously building an AV 1) has their own diverse datasets collected from real-world driving and 2) has almost certainly already handled this issue (skin color is a blessedly minor source of variation in pedestrian detection, compared to the tremendous headaches caused by clothing, costumes, carrying other objects, all of which need to be accounted for...).

The paper is good a reminder of the importance of fine-tuning any pedestrian detection system to fix imbalances in your dataset, and avoid false negatives. But it's not what the title claims.

Likewise. The primary limit on media consumption is, of course, time available to consume it, and piracy doesn't change that.

I think the one interesting place is w.r.t textbooks. Piracy means that I can efficiently pull up multiple textbooks on any given subject, as well as textbooks on any pre-requisite subjects. The size and cost of most textbooks would make that untenable otherwise.

Apart from the standard methods mentioned by other commenters, you can also play CTFs (and read past CTF writeups). The problems in CTFs are not necessarily reflective of the vulnerabilities in modern software, but it's a fun way to learn a lot of the tools and mindset involved in real security research.

> The intent of this article is to sensibilize [sic] the reader that a machine learning approach in not always the best or first solution to solve some object detection problems

This is a good lesson to teach people! But I don't think that this page does so successfully. By far and away the biggest issue is that they are only demonstrating on a single image. You can get almost any method to work on a single image, especially when you're interactively designing the method to work on that specific image. If you want a method that works in unforeseen conditions, you need to think carefully about your assumptions...

* Threshold on 0.5–will that work at night? What if you're driving into the sun and the image is overexposed?

* Dilating canny edges and finding connected components–what if the car is white with black trim, and the dilation merges the license plate CC with another one? What if the car is further away, and the dilation factor merges the entire car into one big blob?

* Filtering the blobs by ratio–what if the car is at an angle (turning, or starting up a hill), so the ratio is off? What if they have bumper stickers that are the same shape, and fool the filter?

While you can make a reasonable license plate detector with traditional methods (see, for example, http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.455...), the advantage of using ML methods is that, if done correctly, they actually have a chance of working in most the situations you encounter (including the assumption-violating ones above), rather than just the ones you can think of while making the system. That lesson is probably just as important as the "don't jump straight to ML" one.

I'm assuming they just made a wallpaper with that text on it and set it as their lockscreen background?

Most of these seem very natural (if you've messed with LEGO before), but I hadn't realized that specific materials choices were the reason that the cone-around-bar construction is a problem (I remember having a terrible time trying to undo that exact configuration). Super interesting deck!

One thing that the presentation only slightly addresses is the standards for overall structural integrity. Official sets seemed to favor a very "light" building style (to make constructions simpler? to save on materials?) that had a reasonably sturdy base (so the set can be picked up without collapsing) but which were vulnerable to lateral attacks or dropping (which would shatter stacked stud-based connections). My experience was that you could typically make models much stronger by adding vertical/diagonal braces via technic pieces.

Summary of Zuckerberg's statements:

* They're still working on Clear History, and think it is worth implementing (Zuckerberg is the one who introduces the topic–the interviewer hasn't asked him about it). It's important that users have a tool like this, even if it makes their experience "worse" (i.e. less personalized).

* Clear History is nontrivial to implement, because Facebook was not built with this functionality in mind. And they want a user interface that is more capable than just a single "delete everything" button, because their research has shown that users frequently want to clear only certain aspects of their data (e.g. data shared to a specific service or app via Facebook), so this adds complexity.

* (In the context of a Facebook subscription service, which they were discussing before Zuckerberg mentioned the Clear History tool)–having Clear History is a prereq for any sort of subscription service, but they want Clear History (and all privacy controls) accessible to all users, not as a privilege for certain users.

I'm reminded of the helpful "How to decide if blockchain is right for your project" flowchart (https://pbs.twimg.com/media/DPFvZuqU8AAeU0C?format=jpg) from @MalwareTechBlog.

Not sure why people are claiming this as malicious. If Apple thought making life inconvenient for linux users was a good idea, this is about the least effective possible way to do that. And it's unclear why Apple would want to do that in the first place.

Seems far more likely that Apple was facing some sort of automated attacks on this particular subdomain (with linux UAs), and a beleaguered admin used this as a quick fix.

Or, even more probably, it's a misconfiguration.

> Safari removed support for DNT so that users are not presented with a misleading and ineffective privacy control that, if anything, only offered additional browser fingerprinting entropy

Predictable, but still a bit depressing to read.

My appreciation to the WebKit team for working so diligently on behalf of their users.

* The given description of Tesla's "Shadow Mode" has been bouncing around for a while, but it's not accurate (see https://twitter.com/greentheonly/status/1096325954375430144 for a description of what Tesla actually does).

* The description of LIDAR having "much more detail than other sensors" is confusing. LIDAR has more detail than radar, but LIDAR is much less detailed than cameras.

* The description of the 2016 accident is not correct. MobilEye provided the EyeQ3 chip on the car, not just the software. "Since then, Tesla added radars for these situations" is flat out wrong; Tesla has always had radar as part of AP, and radar didn't (and won't) prevent that sort of accident (https://twitter.com/artem_zin/status/748623772056428544)

* 'the newest Tesla Autopilot software 9.0 has the largest vision neural network ever trained' wut?

* 'The DARPA Grand Challenge began in 2004'...'the winner, Waymo' wut?

* 'systems based on if-then statements' ?

* 'Comma.ai continues to develop its software, and drivers can buy and install it in their cars' the software is free, it's on Github.

Furthermore, the thesis in the HN title is not supported at all in the post as far as I can tell.

I would recommend reading the paper: https://d4mucfpksywv.cloudfront.net/better-language-models/l...

and the previous paper

https://s3-us-west-2.amazonaws.com/openai-assets/research-co...

It's a transformer, not LSTM, and it's very large but not structured in a particularly unusual way.

You're welcome to check out the samples [https://raw.githubusercontent.com/openai/gpt-2/master/gpt2-s...] and evaluate them for memorization yourself (I haven't found any so far).

(The talking unicorn example on their page is also meant to demonstrate that, no, it's not just memorizing, but I think it's a bit more compelling to check from the raw samples)

Many reactions across here / twitter / reddit seem totally out of proportion. And an odd mix of "stop acting so self-important, this research isn't special so you shouldn't have any qualms about releasing it" and "this research is super important, how dare you not release it".

The strongest counterargument I've seen to OpenAI's decision is that the decision won't end up mattering, because someone else will eventually replicate the work and publish a similar model. But it still seems like a reasonable choice on OpenAI's part–they're warning us that some language model will soon be good enough for malicious use (e.g. large-scale astroturfing/spam), but they're deciding it won't be theirs (and giving the public a chance to prepare).

Absolutely–if you wanted to get an FC layer that satisfies Conv layer constraints, you would need to perform gradient descent subject to those constraints; normal gradient descent won't do that unless the actual optimum is of convolutional form. That's why I said (with respect to expressive power) :)

Practice makes performance at a given task faster, smoother, more automatic, and more robust. The "practice makes permanent" quote elsewhere in the comments is correct in this respect. Practice does _not_ guarantee that anything useful will be learned–this requires some way to measure progress, some measure of deliberate reflection regarding past failures and their causes (which will allow you to propose new behaviors to practice), and, for intellectual tasks, some deeper analysis of the underlying principles (which will allow you to propose better principles to guide future behavior, even in out-of-distribution cases which you have never practiced for).

The article appears to be attributing the difference to genetics alone, which is neither accurate nor a useful line of inquiry. And the article is attempting to argue against a Gladwellized strawman of Ericcson's work, rather than the actual research, which feels intellectually dishonest. Not a fan of this post.

Conv layers are strictly special cases of FC layers (with respect to expressive power).

> For any CONV layer there is an FC layer that implements the same forward function. The weight matrix would be a large matrix that is mostly zero except for at certain blocks (due to local connectivity) where the weights in many of the blocks are equal (due to parameter sharing).

(per http://cs231n.github.io/convolutional-networks/)

In general, if you can make assumptions about your problem and the form the solution takes, you can find a good solution with fewer parameters and fewer data, at the cost of suboptimal performance on problems that violate your assumptions. Conv layers are an example of this.

The dataset is compiled from flickr portraits, and the model was trained to generate sampels from that distribution, so there will definitely be bias towards the sort of people whose portraits end up on flickr. Per the paper:

>The images were crawled from Flickr (thus inheriting all the biases of that website) and automatically aligned and cropped. Only images under permissive licenses were collected. Various automatic filters were used to prune the set, and finally Mechanical Turk allowed us to remove the occasional statues, paintings, or photos of photos.

It's still much more diverse than previous datasets (which used U.S. celebrity photos), but would require some additional work to match the actual world population distribution.

Happy to say I caught this, although I would credit that as much to having taken a compilers course as any JS knowledge.

I'm not sure I understand why it's wrapped in an enclosing function, though; you can reproduce the error with just:

    (() => console.log("Hello"))()

    (() => console.log("World"))()