Hacker News Reader

When was the last time you saw state of the art being applied to a cheap surveillance camera? These aren't video production studios, these are "landlords", which aren't even people, they are housing corporations who spring for the cheapest shit that lets them maximize income and if you don't like it, for you 10,000 other renters willing to put up with their BS to live in NYC.

Facial recognition is an algorithm applied to ordinary picture and video data, it does not require special or even particularly high-quality cameras. In fact, if the footage is too high-quality it actually has to be downscaled before being analyzed; something like a picture of your face taken with a 720p laptop webcam when sitting at the laptop, or a video of your face from a 1080p Ring doorbell camera as you go to ring the bell, would likely be already big enough to require downscaling first.

Yep, typically an algorithm with heavily baked in bias. Works great for specific populations, not so much when it needs to recognize every possible facial structure on the planet - something that's a problem in very few places, except that NY is basically #1 on the list of places where you algorithm has to work for every possible face on the planet, and would have to be tamper proof, which by definition it can't be unless we augment it with IR and lidar to make sure there's a human standing in front of it. And then that would have to still work when it's both 100F and -40F. You know what lets people into your building even more efficiently? A key fob.

Addressing this misconception is the whole point of my original comment, actually. Typically the algorithms used work great on all populations and do not have heavily baked in bias (in the technical sense; they do still have bias in the social justice sense of “automating the detection of criminals, who are disproportionally minorities because society’s bias has disproportionately criminalized minorities”). The field has generally moved away from specific algorithms that analyze facial structure and towards much more effective machine-learning-augmented methods (they got Bitter-Lessoned just like other areas of technology did). Facial recognition has literally already been used to great effect in New York for years already, that’s why it is New York that is leading the way in regulating face rec, because they interact with it the most (apart from maybe Australians at Australian airports).

About the only thing I don’t take issue with in your post is LIDAR/IR to verify it’s a real human and not a piece of paper with a face printed on it. If you’re using face rec to automate entry, yeah, you need to also validate it’s a real human face.

>Typically the algorithms used work great

I hope you let the word "typically" do some stretching before you used it, because it's doing a lot of heavy lifting there. You have no idea what algo was used by each system. Assuming some landlord is going to install the top of the line equipment using the most advanced software with these algos that "work great on all populations" is just so generous that I want to see take advantage of that generosity.

Facial recognition is at best a parlor game. Yes, it can be super helpful for personal use in sorting through your own private library. If facial rec messes up in that situation, no harm is committed. When someone is denied access to their building, harm is committed. When someone is mis-identified and arrest warrants issued on an innocent person, harm is committed. To even begin to think that facial recognition is bullet proof in all instances is just folly.

Respectfully, I do have a much better idea than most people about what is being deployed and used, as I built some of them.

I am not saying it is bulletproof. I am saying it will reliably pull 10 photos of the same person out of 100,000,000 total photos, for any given person, based off a single new photo it has never seen before. That’s the kind of scale and accuracy we’re talking about here, it’s not parlor games. And that 100M was small fries compared to bigger players - who I am sure were into the billions, with even better precision than we had.

I still don’t understand the multiple objections I see relating to “top of the line equipment” or advanced expensive software. The mental model you should have is something like Copilot or GPT3.5. The landlord is just buying access to a nice UI that makes API calls.

I'm with you all the way on your points so far. I think at this point it will be important to refer to a specific product, one that's available to landlords.

I think in the minds of commenters there is an idealised version of "face recognition", so in a way, everyone is arguing past each other.

Go ahead and buy an el-cheapo webcam and test the facial recognition software that is built-in. I challenge you to be able to reliably use that to recognize a face from a block of cheese.

Sure, the high quality algorithms are out there. But no one actually sells the hardware with that stuff built into it, at least not at a price that any sleazy landlord is going to be paying.

If some kind of facial recognition algorithm/model can use worse inputs to "do better than humans" in some task (i.e. try to recognize between twins using a downscaled image from a 720p laptop camera) more often than not focusing in other things different from the face itself or there is some kind of target leakage in the datasets. More often than not those are big reasons why is so difficult to target SOTA paper performance in the real world, be it in face recognition or otherwise.

ML algorithms for FR are not magical and I don't expect them to be always better in all conditions than a human with two eyes (i.e. stereo vision).

I won’t say it’s impossible, but your suggestions do seem unlikely to me. We’re talking about unlabeled data, cropped to just the face, and the state of the art being found in commercial models in the real world rather than in papers competing to pass standard tests in the literature.

(From my direct experience they arguably are pretty close to magical; I think there is just a lot more information in pictures of faces than we thought there was. It’s quite plausible that, since native human machinery was evolved in groups of maximum ~150 distinct faces and was optimized for things like detecting minor emotional cues, our brains are simply throwing a lot of that information out. But that’s not really the point I want to make here; instead, I want to say stuff like “commercial models aren’t trailing research papers; research papers are trailing commercial models”.)

This reasoning (“it needs to be worse to be better”) is indistinguishable from homeopathy.

I meant that it is limited by the size of the input layer. A 2000x2000 pixel image of a face has to be downscaled, because otherwise you’d need 4 million nodes on the input layer and that’s impractical.

Sounds like NYC needs to add more housing supply to take the power away from corporate landlords.

For the landlord situation, what's the additional harm here if the facial recognition system is inaccurate for people of a certain skin color?

From the article, the concern here is that the landlords are "trying to find ways to expedite ways of flushing people out of the building". Presumably this tech could make it easy to identify tenants violating their lease agreement. For example, detecting repeat non-residents using keys to enter the building as a sign of illegal subletting.

But if the face rec system incorrectly identifies someone for potential lease violations, the landlord still needs to go through additional human verification and present proof to a judge to start the eviction process. Especially in NYC, the landlord needs to go to court with a lot of evidence to secure an eviction. So it seems like inaccurate facial recognition would be good for the tenants. If the results are inaccurate enough it eventually defeats the whole purpose of the system.

(this is very different from situations like Madison Square Garden's use of face rec, where a someone might be denied entry unilaterally because of misidentification and have no recourse)

Police/security could be called on a resident every time they walk into the hallway of their building, and landlords would tell them "we've put in a support ticket with our facial recognition vendor, we'll let you know their response in 4-6 weeks".

Or residents could be locked out of the building with no way to get in for weeks at a time.

Just facial recognition wouldn't have a point - landlords are going to use this data for something, and that something is likely to interfere with residents quiet enjoyment of their residence.

Where are these state of the art facial recognition systems deployed? Because as someone with dark skin and a Nest doorbell with facial recognition, I can tell you from experience that Nest still struggles with dark skin.

And if Nest is not deploying state of the art facial recognition then what are the odds that some random landlord is? I'm going to say close to none.

Are there any light-skinned people set up on your doorbell for comparison? Would help to eliminate other possibilities like low light, or the algorithm being generally bad.

Yes, in fact. I have people in my extended family who are fair skinned (i.e. white) and have visited my home. The Nest was almost immediately able to accurately identity them even in sub-optimal lighting conditions. Even though it struggles to identity me except in the most optimal lighting conditions. Despite the fact that it "sees" me a lot more.

It was one of the first things I noticed when I got the doorbell. The difference was stark and obvious.

No, I do agree that Nest’s face rec (and face detection too, for that matter) is pretty awful. I don’t have first-hand knowledge of what Google is actually deploying but I suspect they’re performing so poorly because they’re training the recognition on just the home’s dataset; it is sort of like each home training it’s own LLM just on text written by that home’s inhabitants, while GPT4 gets unlimited resources to train on hundreds of gigabytes of text from all over the web. But it is still a bit of a mystery why Google is allowing Nest’s facial recognition to remain so weak.

Actually, if Nest is representative of what most consumers experience when interacting with facial recognition, then I understand why there’s this common perception that facial recognition is buggy and unreliable. I’m drawing on my experience working on commercial models.

Yes, and "digital cameras will never compete with film," ammirite?

I heard this a lot twenty years ago, and even a little ten years ago.

Sure and state-of-the-art cars will hit the brake before running into another car.

Doesn't mean the car behind you at the red light will brake. Just like the fact that a model can tell twins apart doesn't mean the model the cops used can tell you apart from the actual criminal. Or that a Landlord will have tested the model at all or required it to meet some standard of twin detection.

My usage of “state of the art” was misleading, I apologize. To be more precise I should say “most of the companies offering services in this area are this good and getting better”. This is not like batteries or carbon nanotubes where research is way ahead of production; it is more like LLMs where commercial offerings are significantly ahead of published research offerings.

Can you give some examples of real world successes of this?

Because your (several) comments defending this technology boil down to ‘trust me bro, I’ve seen it in the lab’.

Why do you think they're "defending this technology" when their comment explicitly highlights how this doesn't go against the activist's core point, only a nitpick of a specific argument about "technical progress"?

If I’ve given the impression that I am defending the technology against anything other than claims about its accuracy, I apologize. The fact that I used to work on this technology and now I instead talk about how it harms minorities should give you some idea of how likely I am to defend this technology.

Basically you should just imagine there was a media cycle about the dangers of unregulated trains that go 50mph, do not have any brakes, and need to be regulated because they’re running pedestrians over. And you should read my comments like I’m a pedantic trainwatcher saying “actually these trains go 400mph, can stop on a dime, and need to be regulated because they’re running pedestrians over. Source: I used to be a train driver”.

I have no faith that landlords will be using the state of the art. Especially when the not-state-of-the-art version supports (in a plausibly deniable way) the biases of the landlords. Make no mistake: landlords are absolutely not doing this out of the goodness of their hearts to make their buildings safer.

> The state of the art in facial recognition technology has been superhumanly accurate (e.g. telling apart identical twins better than humans can, finding pictures with significant weight or age differences) on all subgroups for several years now, and the racial disparities in accuracy are minuscule relative to the level of noise.

That may be true of state of the art models that have been demonstrated under some conditions; it continues not to be true, frequently, of models actually adopted in major products, even new ones, in the actual deployment conditions (e.g., the recently adopted CBP One app that US asylum seekers are required to use for appointment scheduling and other purposes, and which requires facial recognition, and disproportionately fails for dark skinned users, who are not uncommon in the population required to use the app.)

Is the state of the art facial recognition technology deployed in real world, scenarios, though? It's one thing to use it in a lab setting with a person sitting still the perfect distance away from the camera, and quite another to operate on hundreds of people at once from a distant ceiling. Not to mention cost (i.e. the landlords don't actually care about accuracy that much, would they really pay a premium?)

To nitpick :) : SOTA is nice but there is often a wide gap between validation dataset and real world usage accuracy. Additionnaly accuracy percentage are often misleading something over 99% sounds great until it's applied thousands of times and you end up with dozens of failure and if those failure are not gracefully handled it can quickly become a nightmare.

And I highly doubt that a "penny-pinching landlord" will install anything close to Sota.

The state of the art in facial recognition technology has been superhumanly accurate

That's SOA and in controlled circumstances you're talking about.

Not real-life FR using cheap cameras, with bad lighting, against shitty lookup databases (with built-in biases). Which is what's at issue, here.

Your apologetic does make for excellent ad copy & pitch deck material, however.

In practice I suspect we’ll see many more misidentifications like the one that happened recently https://www.theguardian.com/us-news/2023/apr/27/california-p...

Have you considered that the activists might be okay with surveillance if the surveillance is fair? That is, if it's not misidentifying people and adds a sense of security and auditability, maybe they would be okay, just as they themselves might install Nest/Ring in their own abode?

Now, I agree that if the argument is "surveillance is inaccurate" then some landlords and landladies might come back and say, "ours is more accurate than trained crowds of professional forensic experts" and it's possible the activists would not want that and just don't want any surveillance, or maybe they are looking for accuracy and the elimination of misattribution.

I have considered that, empirically speaking I’ve seen a few activists do exactly that, and I am definitely not going around believing that activists are all some kind of anti-all-surveillance, defund-the-cops, legalize-crime parody. But also empirically speaking, I have seen that highly accurate systems with no racial bias still disproportionately impact minorities (often because they are employed by institutions like law enforcement or landlords that have racial bias themselves).

I really just think that if activists see two stats like “provably 99.999…% accurate” and “provably impacts minorities more” at the same time, most of them are still going to care about the impact than the accuracy - and I don’t think that caring about the impact is in any way intellectually dishonest, it’s quite moral really.

> The state of the art in facial recognition technology has been superhumanly accurate

On a clear pictures, maybe. On a regular blurry surveillance footage from a gas station camera on a day with bad weather and poor lighting, absolutely not.

https://sitn.hms.harvard.edu/flash/2020/racial-discriminatio...

While I'm sure some algorithm is provably unbiased, it's a little naive to imagine landlords (you know about the landlord special? Load bearing paint and the likes?) caring about what they are running.

I played with AWS rekog, I agree that it is remarkable, but even with a 98 threshold you get occasionally some false positives (but rare), and often some false negatives (I didn't count but I'd say of the magnitude of 0.1-0.5%). That would mean that several days in the year you can't get in because the system mis identifies you.

It’s also an extremely weak argument to say that FR is inaccurate. It’s contingent on an engineering problem that will be solved. Algorithms will get better, HD cameras will become cheaper and more ubiquitous. We are rushing head first into a future where every public and semi-public interaction is recorded and analyzed by functionally omniscient systems, and the temporary inaccuracy of that analysis does nothing to argue against that future.

And then I ask myself, is that really a bad thing? I am far more likely to be the victim of a crime than the perpetrator of one or the victim of inaccurate accusation. The places I’ve lived with pervasive surveillance were also some of the safest, and those who argue against this “oppressive” surveillance rarely are forced to balance that with the actual oppression of crime, or the second-order consequence of the lack of security cameras (that low level crimes will probably go unpunished).

I should also add that going back to a time where surveillance doesn’t exist is not an option. Everyone is walking around with a security camera in their pocket. Very soon those videos will be able to be run through facial recognition tech as well.

> We are rushing head first into a future where every public and semi-public interaction is recorded and analyzed by functionally omniscient systems, and the temporary inaccuracy of that analysis does nothing to argue against that future.

I agree. We need to be fighting the panopticon. The more accurate it is, the more objectionable it is.

Just so everyone knows this is an account that just bullshits elaborately. Everything the account says is just for argument sake.

He explains his point convincingly well.

You should make your point in reply if you disagree with OP, because slandering OP is not changing my pov to yours here.

I'd say the fact that they write-off repeated examples of existing systems that are deployed commercially failing, and then also haven't responded with any real world examples of systems that are "nearly foolproof" as a pretty good indicator of arguing a hypothetical.

No he doesn't. His entire comment history is exactly what he's done here. Decently written but empty and contrarian for the sake of it. Engaging with this person is a waste of time, I was merely providing a public service to others that similarly value their time.