Hacker News Reader

> Rekognition performed so poorly compared to our prototype (which was not state of the art or anything, just good enough for our company’s bespoke conpliance problem). It was just so bad. Overall intersection-over-union scores on our training & acceptance data was really poor, and perhaps worse, the latency of the requests to Rekognition was abysmal and completely unusable for us, especially for images that had a large number of identifiable faces (Rekognition service calls isually timed out at 5 seconds and returned whatever amount of faces it could process).

Criticizing the technology's performance is tangential to the emergence of the technology on public streets, and its adoption by local governments for surveillance. It does not matter what performance it has now; I argue and predict it will improve, and then your statement becomes moot.

The issue is not how well it performs. The issue is allowing it to happen, and without trust from the people.

My fear is: government wants surveillance, and businesses pursue revenue, so there is no incentive among those two to question their actions.

Unfortunately the "cost" of poor performance is sometimes pushed to the public, or worse, subsets of the public. Consider the million-name-long restricted fly list of the last decade -- they were nice juicy profitable projects for corporations but the cost was borne by a subset of mostly innocent people sometimes lost all ability to fly. If the costs are compartmentalized enough, sometimes the general public doesnt care (see: NYC's Stop and Frisk policy.)

Agreed.

I don't see much benefit to pushing back against the technologies themselves; they're here, they're easy to apply, and we're well past the point where trying to ban them will just push them underground. Misapplication of the technologies and erosion of fundamentals of evidence-based action are a real problem.

Contrast the California police using genealogy datasets to narrow the realm of potential "John Doe" serial killers. As court evidence? Shouldn't be admissible. As a tool to focus the search to a subset of the population? Extremely good idea.

Similarly, the face-recognition camera can tell the authorities where to turn their eyes but should never be the be-all and end-all of adjudication; they're too easy to fool (as Maryland discovered when it started auto-ticketing cars that blew through red lights at camera-tagged intersections, only to find a wave of high-schoolers thinking it was hilarious to tape a photo of an unliked teacher's license plate across their own and then blow through an empty intersection at 80 MPH at 2AM).

I agree with you generally. But I differ in the sense that my default belief is that these technologies are over hyped and cannot provide the scale of performance or detail of prediction that could make them effective as oppressive surveillance tools, at least not for a long time.

Mostly they are “surveillance theater” akin to useless (but costly) TSA surveillance. The way it interacts with society is by being vague and scary, not by actually enabling truly scary scales of surveillance.

Another way to look at it is that the government buys these contracts and services and rarely holds the provider accountable, perhaps especially in fields like machine learning where people can more easily make political arguments out of the debate over whether it’s working or not.

Since agencies are happy to drop big money even when the technology does not work then providers like Rekognition actually do not have incentives to invest in making the scale of accurate surveillance larger. Why would they when they’ll make more money yanking an engineer off of a material project like that and slotting them to work on some braindead TED talk sort of demoware that gets them the next big contract?

I worry about the spread of these surveillance contracts more as a taxpayer seeing government money wasted that could help people through much simpler means than I do as a concerned citizen who sees the danger in Orwellian surveillance.

These surveillance theater and immature or improperly applied technologies do more than just waste tax dollars. For decades the courts allowed flawed hair analysis to be used as evidence. When it could result in innocent people being sent to prison, a fuss must be made that's stronger than just worrying about what it will cost.

That’s a fair point, which makes me feel even more convinced that we should be more focused on whether the technology actually works than whether it’s Orwellian.

Using Rekognition for this would be like hiring a fraudster construction company to build a local bridge. The bridge could collapse, nobody did publicly verifiable due diligence, and nobody would be held accountable. This could be the automated surveillance equivalent.

>my default belief is that these technologies are over hyped and cannot provide the scale of performance or detail of prediction that could make them effective as oppressive surveillance tools, at least not for a long time.

you've never heard of the chinese police state's facial surveillance which they're using to throw countless uighurs in jail for stuff like jaywalking, i suppose?

>as a concerned citizen who sees the danger in Orwellian surveillance.

you don't see the danger because your paycheck relies on you not seeing it. the rest of us are not so willfully blind.

But I do see danger. Both in the misapplication of this unproven tech, and in the opportunity cost of wasting public funds on it.

then why work on it when it's bound for abuse either way?

It seems misguided to dismiss the technology itself, rather than scrutinizing specific uses. It reminds me of David Deutsch's book The Beginning of Infinity, which has an extended anecdote about a scientist who deeply objected to the development of color TVs because "people didn't need them" and he could only think of examples where bad actors would use them. Fast forward several decades, and color TVs are used for all sorts of important outcomes, like life-saving medical imaging technology, or sharing video of an important life moment with someone who is too sick to travel.

You could imagine similar things for face recognition. An Alzheimer's patient who has a large digital photo library of family members and events, and would like some information retrieval tool to help search for a specific memory, family member, etc. Or someone who takes stock photographs and wants to photoshop bystanders out of the photos so that nobody's image rights are infringed when submitting the stock photo for publication-- might want face detection to aid in the person-removal task.

I agree completely with scrutinizing motives and refusing to work on tech when the goal is unethical. But I don't agree this allows us to point at a generic, amoral technology in and of itself and say it's conceptually unethical for existing. Maybe there would be some extreme or contrived cases when I would agree to that, but even this face recognition stuff is very far from it.

How many technologies can we name where this question does not apply?

The answer is always the same: "Because the potential for benefit outweighs the potential for abuse."

It does, however, suggest that whether Amazon's doing it today is irrelevant, if any org with the resources of mlthoughts2018's org could cobble together a system that performs sufficiently to their needs with open-source solutions and some locally-grown or publicly-sourced training data.

This barn is devoid of horses and is getting chilly from all the airflow.

> backed by a team surely comprised of at least hundreds of top Amazon engineers, was so slow and inaccurate.

You'd be surprised. Many many AWS services are launched by a single 2-pizza team (8-10 developers). It's in Amazon's DNA to act like a startup, ship something quickly, collect feedback, then iterate, and scale.

Rekognition will get much better in time.

Also, from what I’ve been able to deduce, the AWS services are all just built on top of EC2 [un]modified open-source software until they reach a level of complexity where they roll their own, which leads to lots of issues that can take years to fix. Some things are telling, like banning two consecutive hyphens because something somewhere in that pipeline is shelling out.

Sidebar: I've never heard of teams measured in terms of "pizza count," and now I can't un-think it. Good metric. :)

It's been reported that Bezos uses a two pizza rule for meetings.

8-10 people and only two pizzas??? Madness

Also, for the record, AWS is hardly top of the industry in terms of talent. It's not Google or Facebook. I've never heard of an ML researcher joining Amazon when they could have joined Google/FB/Apple.

I've known plenty of people that have left Google/FB/Apple for Amazon. And I know plenty of people who have left Amazon for Google/FB/Apple. The biggest tech companies are quite incestuous. Some people do thrive in one environment, some in others. They all pay top dollar.

It could happen both way. Noted that Google is much more saturate for its ML talents, so unless you are super top-notch, some people will take it to Amazon for bigger impact. I personally know a few people, who is competitive in the industry, having Google offer/leaving Google for Amazon.

No doubt, the flow is usually Amazon -> Google/Facebook, because for mid-level engineer/scientist, money talks :), and Google/FB generally have much open/better publication policy, while Amazon is neurotically secretive. Apple is not on the picture, I think the internal engineer culture is very problematic(no common codebase, severe NIH syndrome), except for the money, it is subpar place to grow.

Disclaimer: Work in a ML related AWS team.

I wonder how many innocent people are going to be harmed due to poor facial recognition matches...

Indeed! I recently tried Rekognition's "celebrity detection" and was surprised at the no. of false positives - "doppelgänger detection" would've been more like it.

Some examples: https://medium.com/@deepak.gulati/trying-amazon-rekognition-...

That said, it could flag faces for manual screening. Or cross reference other data (License plates, Geo data from phone) to boost confidence. Also, once they get more data as part of some of these deals they'll get better.

Your post makes me think of this article which pretty comprehensively reviews it

https://cloudacademy.com/blog/google-vision-vs-amazon-rekogn...

The scenario-based costs table at the bottom is particularly interesting. If you need to run the detectors on all incoming images of a site ingesting, say, ~10 million images per week, you're looking at nearly $15,000/week, or close to $800K/year with AWS, and nearly $1.2MM/year with Google, and either way you'll still probably need to employ some type of backend engineering team to maintain the wrappers for calling it and manipulating the response data.

We were able to solve this in-house using only about 15 8-core servers (we used a few general GPU servers for training, but found we only needed CPU machines for runtime, and did not have the poor latency of the AWS calls) with quite a lot of redundancy for traffic spikes and a pretty easy deployment system if we needed to add or remove nodes, and it was only one of several dozen different machine learning projects going on, so whatever portion of the total cost of operation could be allocated to paying the salary of a machine learning team, it would further be amortized by work on many projects. For example, amortizing the cost of GPU machines for model training across a variety of projects.

interesting data point on its performance but the reason decision makers might be cloud happy is amazon has the resources to scale this to thousands of RPS and develop the backend codebase up to hundreds of engineer contributions for many years, even as key talent comes and goes. The same might not be true of your prototype especially if the core competency of your business is sales or legal compliance not distributed computing platforms.

Our team was just a general machine learning team with no expertise in compliance issues, and only one or two of us had experience in face detection. We only cared about providing a web service that gave good face detection performance (the downstream consumer was the team that had to then take the face detection output and do business logic with it for compliance).

> amazon has the resources to scale this to thousands of RPS

Actually, this part is mostly straight forward. Not easy, but straight forward, and Amazon already has this part solved for most things. Our ML team did have expertise in making very large-scale services (our company operated a large ecommerce web store, so our traffic was actually very high, top 500 Alexa rank), but our concern over Amazon had nothing to do with whether it could handle enough requests per second.

But this also misses the point. The latency for a single request was too poor for a variety of natural images (crowd scenes, public streets, etc., -- exactly the types of images relevant for many surveillance applications). It was not a throughput issue, I guess unless you expect consumer teams to do something like chopping images up into sub-images, replacing one Rekognition request with a bunch of sub-image requests, and then post-processing to stitch the results together and account for error introduced by the choice of where to split the original image. Yikes.

It's not an issue of "scaling it up" -- it's an issue that the fundamental machine learning algorithm underpinning it suffered too high latency, likely because it was a deep neural network combined with a bunch of post-processing gadgets that are conceptually not good enough for bespoke use cases -- things which engineers on a centralized platform have little incentive to care about.

The real question is whether Amazon has incentives to make this part better, and if so, how? Will a general one-size-fits-all face detector work for customers? Or will they need application-by-application customer-specific tailored models (and if so, why aren't customers savvy enough to realize that the all-in costs to build and operate it in-house are so demonstrably cheaper than Amazon, and might be the only way to meet specific performance goals)?

I hear this so often, that Amazon can "just scale it up" -- but I think if you really work in face detection and have studied the problem deeply, you know this is not right. There's a whole different bucket of problems and issues and cost-effectiveness trade-offs going on.

As it get customers, it will likely get more and more resources thrown at it and get better and better.

Even if the tech may not be useful right now, it probably will be in 5 years, and unless something is done now, another large slice of the little bit of privacy that remains will be gone.

I think this is unlikely because of the nature of the problem. It's kind of like when Paul Graham said consulting is fundamentally not scalable. Throwing more people at a problem like this one only works if you have dedicated consultants for every bespoke business case, because there's no one-size-fits-all model, and the way that performance limitations interacts with which kind of face detection models you can use is very complicated.

Companies would end up paying huge premia to Amazon essentially to have Amazon build in-house teams that would be equivalent, only much more expensive, than the companies own in-house team just building an in-house solution.

With some forms of consulting this can make sense, because the consulting services are temporary and should set the company in a position such that for the long run, the company can maintain the solution cost-effectively.

But for something like face detection, assuming it's a pivotal service you'll need on an on-going basis, this would become the worst sort of vendor lock-in imaginable, combined with the fact that if you don't appear to be creating growth for Amazon, you're liable to be deprioritized at any point, and at the mercy of Amazon's choices regarding when and how to address your bespoke need.

For these specialized machine learning services it doesn't fit the same kind of commodity model that AWS infrastructure uses, though that's currently how Amazon is pursuing it.

Don't get me wrong, Amazon is very good at making money and convincing nervous management to buy their brand. I suspect they will find ways to sell these services despite the services not being cost-effective or customizeable for bespoke customer situations.

But when I point out that Amazon's machine learning services don't offer cost-effective performance, it's very different from a "just throw more devs at it" kind of problem, fundamentally.