Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Deep-Fried Data

idlewords.com

Post reply on HN
Print view View original

141–149 of 149 posts

Re: Deep-Fried Data

#141

Post by yummyfajitas »

eridius wrote: post #139
yummyfajitas wrote: post #138

Earlier quoted context omitted.

Your function r = r(pp, red shirts, race of offender, etc) exists. A model of the form a x r + b x something_else + ... will detect the bias you've described, assuming of course the biasing variable is either present or redundantly encoded in the data set. We've now established the existence of a statistical model which can detect this bias. Now, any other model which is capable of expressing your specific r(p) can d…

Regarding your last paragraph, there's two different angles here. The "machine learning is racist" angle I think is quite valid, but covers a different topic than what we've been discussing here. To be more specific, there's two different ways in which we can have racist models: 1. The algorithm is biased in a way that reflects reality but does not reflect how we wish it to behave. This is the "machine learning is ra…

Here's my point.

(1) is only possible if your data provides access to the biasing variable, perhaps via redundant encoding. This is the standard critique folks make.

As per (1), the biasing variable is available. Now if the algorithm is expressive enough to describe the functional form of the bias (e.g. the bias is quadratic, and the model includes quadratic terms), it will fix that bias.

You're right that there are lots of hidden variables that we can't use in a predictor. Murderous intent and mafia membership are also not available as predictive factors. You could build a more accurate model if you had that data. So what?

Re: Deep-Fried Data

#142

Post by eridius »

yummyfajitas wrote: post #141
eridius wrote: post #139

Earlier quoted context omitted.

Regarding your last paragraph, there's two different angles here. The "machine learning is racist" angle I think is quite valid, but covers a different topic than what we've been discussing here. To be more specific, there's two different ways in which we can have racist models: 1. The algorithm is biased in a way that reflects reality but does not reflect how we wish it to behave. This is the "machine learning is ra…

Here's my point. (1) is only possible if your data provides access to the biasing variable, perhaps via redundant encoding. This is the standard critique folks make. As per (1), the biasing variable is available. Now if the algorithm is expressive enough to describe the functional form of the bias (e.g. the bias is quadratic, and the model includes quadratic terms), it will fix that bias. You're right that there are…

The problem with (2) isn't just that your model isn't as precise as it could be, it's that your model may be inadvertently biased because all of the data that it was fed was biased. This comment (https://news.ycombinator.com/item?id=12625917) gives a good example of that one. No amount of expressivity in the algorithm will account for the fact that the Friendface model (read the comment) was trained on a predominately white userbase versus FaceSpace's model which is trained on a predominately urban black userbase.

Re: Deep-Fried Data

#143

Post by keyle »

odbol_ wrote: post #127
keyle wrote: post #44

Earlier quoted context omitted.

Same here. I'm battling with this thought a lot. Beyond jobs, I think there should be communities of developers, designers, producers, writers, getting together and figuring out this stuff. And I don't mean open source projects. Let's group together smart people wanting to make a difference and have a hit list of things we (people) actually need. A group that would organise people into mission driven development. I'm…

I think the First Things First manifesto is a good start: http://firstthingsfirst2014.org

That is great thank you.

Re: Deep-Fried Data

#144

Post by srean »

yummyfajitas wrote: post #121
srean wrote: post #120

Earlier quoted context omitted.

> But if your model is sufficiently expressive you don't need to explicitly build or model the corruption process This is the claim that I am having trouble with. Say I have two random variable X,Y with some joint distribution. If a corruption process can mess with the samples drawn from it, I cannot see how it could possibly recover either the joint or the conditional. Are you saying that the corruption is benign li…

So we both agree that if the bias is linear, and your model is linear, you capture it. Similarly if the model involves interaction (score x is_black), and you include linear interaction terms, you'll also capture it. Now the question arises; what if things are more complex? In real life they always are; both your biasing factor and the rest of the model. So we've cooked up all sorts of fun models like SVMs, random fo…

That makes it clearer. I got thrown off by the claim that a standard algorithm will be able to de-bias if no de-biasing machinery has been built into it. BTW the machinery may be implicit in the choice of the model.

Simple toy example: say Y is a threshold function of X + high variance noise. I draw samples from this and scale down all y_i's that exceed the (unknown) threshold. In other words my corruption process is dependent on X. We can make it depend on Y too. These would require explicit modeling. Just throwing a uniformly rich class of P(X,Y) wont by itself fix this. We have to carve that space of P(X,Y) with the knowledge of possible corruption process to get a good model of the behavior before the corruption is applied.

BTW we have gone way off tangent, but that was a good conversation.

Re: Deep-Fried Data

#145

Post by srean »

yummyfajitas wrote: post #138
eridius wrote: post #137

Earlier quoted context omitted.

No, I'm claiming that P(crime detected) != r(pp). More police in an area typically means more crime is detected, but that's not the only factor. If you have two areas with identical police presences and identical actual crime rates (as opposed to reported crime rates), the rate of crime detection (as measured by arrests and whatnot) may be higher in one area due to other factors such as racial bias (not just racial p…

Your function r = r(pp, red shirts, race of offender, etc) exists. A model of the form a x r + b x something_else + ... will detect the bias you've described, assuming of course the biasing variable is either present or redundantly encoded in the data set. We've now established the existence of a statistical model which can detect this bias. Now, any other model which is capable of expressing your specific r(p) can d…

>Your function r = r(pp, red shirts, race of offender, etc) exists. A model of the form a x r + b x something_else + ... will detect the bias you've described, assuming of course the biasing variable is either present or redundantly encoded in the data set.

No no no. Had to respond to this because this such a common confusion (not to say that you personally have this).

That such a model exists within the class of models being used says absolutely nothing about whether the statistical/ML algorithm will find it with any degree of confidence from a sample. The science is still grappling with the question of how to do model selection. There are two, sort of, equivalent class of methods, regularization (this can be a regularization over the dependency structure too, not just a simple penalty) and prior. Its only when you get those right that you have decent chance of estimating well, from reasonable amount of data.

Short answer: universal approximation property of a class of models says nothing about learnability.

Re: Deep-Fried Data

#146

Post by sangnoir »

coldtea wrote: post #122
sangnoir wrote: post #99

Earlier quoted context omitted.

> That software (desktop Linux/UNIX) comes for free, and yet, only 1% or less (from browser stats of major traffic points) seem to opt to use it as their desktop. >Is this the kind of difference the parent was describing? Cancer kills 171.2 per 100,000[1]. So by your metrics, the Linux desktop folks make a bigger difference than curing cancer as 1% > 0.1712% 1. Cancer mortility. see https://www.cancer.gov/about-cance…

> Cancer kills 171.2 per 100,000[1]. So by your metrics, the Linux desktop folks make a bigger difference than curing cancer as 1% > 0.1712% This is the kind of illogical result stemming from only reasoning half-way. First of all, the cure for cancer wouldn't affect only the ones that die but also the ones that don't but do suffer complications from current treatment, from going broke from paying for therapy/losing t…

>>Cancer kills 171.2 per 100,000[1]. So by your metrics, the Linux desktop folks make a bigger difference than curing cancer as 1% > 0.1712%

> This is the kind of illogical result stemming from only reasoning half-way.

I'm glad you saw the flaws in your reasoning - the key phrase in that paragraph was "by your metrics".

> Second, ever considered the kind and magnitude of impact?

My point exactly!! It's not just about numbers/proportion game where you get to say "1% or less [desktop Linux usage] is not a big difference" since it means the world to those who depend on it, for example those who cannot afford Windows licenses or an Apple computer or who find non-free software unconscionable.

Re: Deep-Fried Data

#148

Post by joepie91_ »

qwertyuiop924 wrote: post #25

>the Internet is a shopping mall. There are two big anchor stores, Facebook and Google, at either end. There’s an Apple store in the middle, along with a Sharper Image where they are trying to sell us the Internet of Things. A couple of punk kids hang out in the food court, but they don't really make trouble. This mall is well-policed and has security cameras everywhere. And you guys are the bookmobile in the parking…

As somebody who has been involved in the operation of various things in the "shady back alleys" category, I can assure you that they do not have an established place on the internet anymore. It's a constant battle of keeping things online, with pretty much every service provider involved looking for a way to kick you out. You may not see this as an end user, but it is exceedingly difficult to host something that goes outside of the norm.

The idea that the internet is a city might have been true 10 years ago, but it is definitely not true anymore now. The default response to "I need an X" is "just do it on Facebook", and there are entire swaths of content that just don't have a place anymore on the web, "thanks" to ever-increasing enforcement of arbitrary moral guidelines and growing monoculture.

If the internet were to be described as a city, it'd be a gentrified city where most of the artists have long been chased away by ever-increasing rents.

Re: Deep-Fried Data

#149

Post by dkersten »

gabrielgoh wrote: post #20
absherwin wrote: post #7

Earlier quoted context omitted.

Unsupervised refers to whether or not the dataset is being trained against anything. Think about the difference between: How many people will view this webpage? Divide these pages into 20 clusters? The first is supervised. The second isn't. Deep learning refers to a particular type of a particular learning technique: Specifically a neural network that has many hidden (intermediate) layers. Deep learning can be used f…

I agree with your sentiment, it feels out of place because deep learning, AI and big data are buzzwords, but unsupervised learning is a rather technical term in machine learning referring to a very specific class of problems.

Are they really buzzwords? To me, they have rather particular meanings (although I guess others may feel differently):

Deep learning: a particular type of artificial neural network with many hidden layers (and the associated tech to make this work/trainable)

AI: The field of computer science which aims to make computers smarter. Like most fields, there is much overlap with others, for example, statistics.

Big Data: A buzzword. About the best definition I can find is anything which has the 3 V's: Volume, Velocity & Variety. In general, outside its use as a buzzword, I think big data is generally thought of as "when you need a distributed system to process your data", be it because of volume, velocity or variety.

Supervised and unsupervised learning: whether or not you require example data for training

Machine learning: some people say its the subset of AI that deals with statistical methods, other people say its just another word for AI.

Post reply on HN