You cannot eliminate bias. You can only compensate for it by illuminating more options. Otherwise the bias "elimination" is subject to bias. E.g. If you avoid a subject when teaching someone, it becomes a weakness in their understanding, and can fall into an overcompensation bias. Furthermore, who decides what counts as a negative bias that should be eliminated? That strikes me as the kind of thing we should be having discussion on and not deciding for other people without their consent. Give people more opportunities to understand, not fewer opportunities to learn.
Google: Makes a promotional video in which they directly ask people to join the conversation about bias. You, an intellectual: "That strikes me as the kind of thing we should be having discussion on and not deciding for other people without their consent." A yes, I see the word understander has entered the room.
So who decides what's biased? Does "equal inclusion" mean the results are unbiased? What if the unbiased view of those engineers overlooked by policy makers within Google isn't actually unbiased? Put simply, who will guard the guardians?
But, what if I am trying to find the hateful stuff because I am trying to see what other people are saying? Doesn't matter if they are morally wrong or right, it should still be easy to find
This is just replacing one bias with another. How about just letting every person decide on what they want to see rather than automatically deciding based on some "offensiveness" interpretation that others are doing for me. Furthermore, how about allowing some negative speech? Yes, the internet can be a bit of a cesspool at times, but without adversity we grow no stronger against it. We can't always rely on something being a safespace tailored to our needs. Please stop being silly Google..
But could those two things go together? If they tailor the user experience individually, then the result would be similar to a Facebook feed: an echochamber of similar ideas. But if we want to allow negative speech (relative to the user), then they would likely not want to see it, which would go against the goal of tailoring experiences individually. However, I still think the individual user should decide what they want to watch and that Google shouldn't participate in censorship.
Something that should be noticed is that just because most Physicists in the past were men, doesn't mean there's a bias. It's just a fact. Same goes for females and great teachers, shaping people's lives.
William Lardner The issue appears when the machine is used with this bias, like categorising photos, or if you ask it to show you pictures of physicists. It might not even recognise a female physicist, which is a mistake in the program. A bias might have a good reason to be there, but that doesn't mean it should still have influence.
Never thought about machine learning and human bias. Always thought it will not affect the results. But we are designed to see the world from our own eyes, experiences. Why will our code be any different.
Bias can actually be a good tool for a computer. A physicist doesn't look like anything, and you would want a computer to understand that. But, if for some reason you need a computer to be able to pick the most likely physicist out of a lineup, then it would need that bias to form an educated guess. A computer should not be free from bias, it should just know when to use it.
This video starts out with a great description of human bias and then proceeds to say that, to resolve this problem, Google has selected the biases that it thinks should be enshrined... that's not a lack of bias. That's enforced and canonized bias.
This video is relatively educational and presents clean information, then I see people force their opinions about adjacent subjects in the comments. How appropriate.
If those biases happen to reflect the truth, are you not suppressing the truth by artificially injecting a bias of your own? It's like stereotypes; on an individual level they are socially inappropriate and misguided, but they're often reflective of some reality at the group level. Should that reality be suppressed?
If you're training your model on what a physicist looks like, and you use a training dataset of past physicists, you are training a model on what past physicists looked like. There is no bias there. The problem is training data that does not match the question that is trying to be solved. Certain people are using these problems as a justification for interjecting *actual* bias through human intervention. It's dishonest and anti-intellectual. Honestly, this is machine learning 101. It's just issues of overfitting and underfitting. Poor data sets. I guess they've started to call it "bias" so that they can claim a moral high ground while controlling outcomes to their liking. That's what they've been doing by the way.
@@GrantGryczan The bias of the individual or individuals making the changes to the results. Let's say I take a poll in my town: "What's your favourite music genre?" If the result of that poll is that most people in my town are fans of acid jazz then that is the result of the poll. If I think acid jazz is terrible and more people should discover the wonderful music of Justin Bieber then I could change the results to give justin Bieber more exposure and (hopefully) get more people listening to good music instead of that awful acid jazz that so popular. What I have just done is introduced my own bias into the results. This is exactly what Google is doing while claiming they are "un-biasing" the results.
@@ZoomahZoomah Refer to the other comment chain: Google isn't doing a thing, and the bias the video refers to has nothing to do with sample ratio accuracy.
It appears that google is deciding what is not biased. How are people at google able to be sure that they are not introducing their own bias into this process?
The very act of setting up the system introduces biases no matter how much they try or say they try to avoid it. By trying to eliminate biases, they introduce biases. Who are they to decide what is a good or bad bias?
@@reverendcaptain They don't decide. They let the machine do it, and then they let the users moderate. Again, this is demonstrated in the video. The employees are not a part of the process but to maintain the machine, and they do not give the machine input. By not giving the machine input, they are not introducing any bias. The bias they refer to in the video is created by the users, and the solution to resolve this bias is also by the users. The employees in particular do not introduce nor try to eliminate bias.
@Grant Gryczan Based on this video Google provides tools for users to remove opinions or information they don't like. This means that certain groups of people can manipulate what kind of information of them or their interest is available for public. This does not promise good for any kind of minority views. This leads to exact kind of bias they wanted to eliminate on the video. (Caricatured example: Take male scientist bias. If majority of viewers did not like - for whatever reason - female scientists, they could remove female scientists entirely appearing on the search results.)
@Grant Gryczan Did you delete your reply because I can't see it? Did you realise that my female scientist example was in parenthesis because it is ridiculous example and you need to replace it with some other interest group of your choice. I guess you can imagine examples of interest groups where majority of the group does not want minority of the same or opposing group gain visibility.
First step by AI to take control from humans. By appearing unbaised & pointing human biases. E.g. -Divided men & women then become the unbiased judge to dictate! A human system is to be controlled by humans not machine!!
This is the scariest thing in the entire world....What gives Google the authority to decide what is and isn't negative, bias or hate speech? Alphabet a.k.a Big Brother/Skynet.
Luckily this is still a free country and we still have the vote of the dollar and where we put our resources and funding. I deleted my FB account I had for ten years because I still had that freedom. WE give them that authority, and we can also take it away
"Report inappropriate content" is also biased. This means it only counts those people who think content is offensive but doesn't count those who think that the content is not offensive.
so how do you eliminate the human bias that controls the moderation of the machines human bias? doesn't seem to be much help the "limiting of offensive results" only removed "offensive" opinions that google doesn't agree with, either manually or through new human bias influenced machine learning. opinions like that of the man who google recently fired for questioning google's current stance on workplace sexism. even if you agree with google for this example, there could be anything that google finds offensive that you don't. if the only information available is the information not censored by google, whether or not you think that the results would be in your personal favor, the control over what opinions people have access to should be the right of no person or organization.
Hi. We're Google. We support Facism under the guise of compassion. Don't worry though we're got a BRILLIANT marketing department. We'll make it all feel like its all a nice warm bath.
Went from an informative lesson on machine learning to focusing on what Google considers appropriate. Why are you training ai to learn based on your own bias?
The whole idea of freedom of speech is to be able to say almost anything as we seek to discover the truth, so yes, even if it offends someone. Someone who is in denial may easily consider the antithesis of the position they hold to be offensive, even if the opposing view is true.
Bias many times begin with those who think others are bias! However, it does not occur to them that their "Politically Correct" ideology itself biased, which makes them to see other views as biased if it does not confirm with theirs. And the political blame game begins.
If most physicists are male then it isn't biased for a computer to generate more male physicists. That's not what bias means. If something makes decisions based solely on facts then that something isn't biased. Something being offensive to a lot of people doesn't make it a bias.
I was with you until your last sentence. How does any of that logic show that the video is biased? On that note, what is the video even biased toward, and how? Edit: They edited the original comment, so never mind this one.
@@GrantGryczan I might be wrong but the video seems to be biased towards a certain branch of identity politics by assuming that representing everyone is equivalent to not being biased.
@@sel2230 This video doesn't even reference identity bias. Did I miss something? As far as I know, it only talks about bias in AI recognition. That's why I agree with the rest of your comment; everyone is going around the comments and bringing politics into this video while it has nothing to do with the biases at hand.
@@GrantGryczan you might not be familiar with Google's affirmative action policies. That and the wrong message this video portrays makes me think that this video is biased. However, you might be right. Maybe the video is wrong because of other reasons.
@@sel2230 I'm not sure you realize, but this video isn't trying to portray a message. It's just trying to show people how their AI works, particularly to resolve automatic inaccuracies. This is an objective informational video.
If you're able to localize the recognition problem you can greatly intros the accuracy of your models by weighting that localization heavily. You don't necessarily want to include everyone in the solution step. You could even train local networks.
"we've been working to prevent that technology from perpetuating *negative* human bias". Right. So you'll be working to PREVENT *negative* bias, but NOT *positive* bias... Who gets to decide whether a particular bias is, on the whole, negative or positive? And surely you'll be tempted to ENFORCE *positive* bias to socially engineer your "positive" ideals. Any bias can be rationalised as a *positive* bias, so the use of this qualifier is legitimately frightening. You purposefully and publicly leave the door open to manipulate your machine algorithms, and by extension your users, based on what "Google" thinks is *positive*. We'll get an intersectional affirmative action AI from Google soon, while Google will claim publicly that it won't have a bias. We can see the precursor for that on TH-cam already. I believe that is actually *evil.*
Terrence Koeman “well get an intersectional affirmative action AI from Google soon, while google will claim publicly that it won’t have bias.” - Terrence Koeman
First step by AI to take control from humans. By appearing unbaised & pointing human biases. E.g. -Divided men & women then become the unbiased judge to dictate! A human system is to be controlled by humans not machine!!
@@cafeta Whatever that is is not relevant to the video. Again, they explicitly described systems implemented for users to be able to resolve the network biases. If you're just going to ignore those along with the point of the video then I'm just going to ignore you, because what you're saying to this video is not relevant.
Good video. I'm putting together a laboratory informatics summit which has a strong focus on machine learning - and I wonder how much this effects things like new drug discovery or data analysis.
To try and modify statistics in order to generalize them to be false is in fact, biased. Commanding an A.I. system to collect available data in correlation to key words instructed by a user resulting in correct, specific and factual data is not biased. Statistics are averaged for practically based on questions that are variable such as "What does a shoe look like?". To use the reasoning that less images of women physicists appearing from image search results as a bias is false when the factual statistics are only being relayed by the A.I. system because they are in fact less common; they will be less likely to show up due to practically, not bias. To alter this information would make you biased. You're reasoning in multiple regards, including the shoe result example, are false and hypocritical.
This is not about result ratios. It's about recognition. They never said or implied image search results for "physicists" should return equal male and female. They just said the AI should be able to recognize both a male and a female physicist. To be able to recognize the latter, you need to unbias the data so there are fair samples of both.
It's the problem of inference vs. prediction. Statistical inference might show that women are less likely to be a physicist. And it might be revealing a *problem* in our society. For example, 100 years ago, you can hardly find any Chinese physicists, but would you use that data to make a prediction that a Chinese person is not likely to be a Physicist? This prediction would be laughable today, but if AI existed 100 years ago, it would have made that prediction. The problem is that AI look for patterns, not theories. And that is the risk in believing that AI/ML is objective.
Judging from the comment section, it seems too many random people with no idea of machine learning let alone weights and biases and how they are incorporated in learning processes have stumbled upon the video. Not every video is meant for your poltical opinions people.
@@claytonwoodcock6942 They don't do that. They let users report results as inappropriate, which are thus automatically removed. It has nothing to do with political views.
@@claytonwoodcock6942 This system is not one of censorship; it's a system of AI clarification. Because it was not designed to censor, trying to use it to censor inevitably won't work very well. AIs don't know what's related or unrelated to what. So users correct them when their faulty automatic predictions...are faulty. That's all this is. I don't know why you're associating that with censorship. Removing opinions people don't like wouldn't be very effective. For example, no one is going to go and search "religion" and then report all the results related to Islam as not relevant just because they don't like Islam. It wouldn't be useful anyway, not only because it would happen on all sides of the topic (not just Islam), but also since the search term "religion" appears so often with the term "Islam". The AI would just retrain itself to associate the two. Plus, you'd have to go through thousands to billions of results to do this, since they are so strongly associated already, which is never going to happen. This system only works for relatively small exceptions to accuracy (as is intended), where the neural network doesn't have to change so many connections to correct the biases. All of these factors would apply to any opinionated search topic, not just religion and Islam.
"Offensive", "hateful", "misleading" and "representative" are all ideas completely constructed out of human bias. Doubling down on arbitrary bias does not remove preceding bias, it just enforces yours instead of someone else's. The irony of you identifying an issue and then embodying it while claiming to be mitigating it is hysterical IMO.
everything human/emotional aspect is subjective, but if your subjective beliefs, stereotypes lead to tangible and measureable affects on the real world that is problematic. So you might be right in saying that it is subjective to say statement X is racist, but even so if statement X influences or leads to mass detention/genocide of race Y members, then wouldn't that be problematic regardless of who determines what is racist.
And then again there are biases of the people who flag the search results shown at the end of the video. So in reality what we need to remember is that what we imperfect humans with biases create will also be imperfect & be biased.
You have a bias to only look at the most relevant comments, which has a bias to be recent comments. Your thought about the comments is biased and wrong.
But isn't that bias in it self? What you consider offensive, others do not! Even today what we consider offensive in our own society is changing. Is it not?
@@johnnybadmen3473 I like to resolve practical misinformation. I check for new comments whenever I get a reply notification here. If you have reason to continue disagreeing with me, feel free to reply with said reason to form a logical argument.
First step by AI to take control from humans. By appearing unbaised & pointing human biases. E.g. -Divided men & women then become the unbiased judge to dictate! A human system is to be controlled by humans not machine!!
That's fine, because there are more male criminals then female criminals and it's just a search. If you are trying to predict whether someone is a criminal however then you have to account for the bias and make sure that it doesn't just clear every women or over accuse men.
It's based on your personal preferences (e.g what you mostly click on.) You are training the AI for your recommendations. I am training the AI for my own. Welcome to machine learning.
Hey, Google: The example code @ 0:49 is missing a second = in the first line to establish a Boolean for the while condition. (Can I get a job interview now?)
Hey Google, Are you suggesting that if humans are to improve themselves, they should be more like brainless machines who should be TOLD what to think and how to feel about things?
- 95% of physicists are men - The problem that would occur : Google results show 100% men. - The desired outcome : Google results show 95% men physicists. - The BIASED google answer to the problem : Results show 50% men physicists.
It was very, very, very... very obviously an exemple. Google used shoes, I used physicists, the reason I chose a gender-related data is because that's the kind of data they've wrongly altered because of their own bias. Which was the point of my original comment. ------------------- But if you're actually curious : news.cornell.edu/stories/2007/04/where-are-all-women-physics [...]The low numbers of women in physics, she said, are especially shocking: Women in the United States hold less than 5 percent of full professor positions and make up only 22 percent of the undergraduate majors and 16 percent of the doctoral candidates. At Cornell, women comprise 17 percent of physics graduate students.[...]
Another thing to keep in mind is that there is a difference between *bias* and *context* . Who were those people who were asked about the shoes? When was it? Where was it? Removing all traces of potential bias can lead to irrelevant search results etc. And the case with neural networks not recognizing faces with different looks is not really connected to bias in my opinion. No wonder that a network that was trained with millions of western looking images is bad at detecting an aborigine from the deepest outback. That's not a racist or biased computer - it just doesn't know better. Adding the image to the face database later on is important of course, but all in all it is still an improbable outcome to find an aborigine in comparison to western looking faces. That's exactly where context comes in - when a face search is performed in Australia, then the expected results should very much lean towards aborigines as well (Sorry aborigines - you're my scapegoat today^^), but when I - here in central Europe - want to recognize a face, it is not wrong or racist to assume European looking faces usually.
Every data itself shows a biased idea to human brain(because it was created by human logics itself)...So as far I understand I think we can neglect the bias almost in all cases(but still there are chances of failure)... :)I found this satisfying;)
It went from cool and informative to "OH we are using our understanding of bias to improve censorship". I think you could have stopped the video 30 sec earlier and people would have been happy, but at least your honest and gave the real reason you are developing this. I mean How about this: have a video talking about how human bias effects data collection on the level of science, bias in the results, and how machine learning could be effected. This is far more useful and interesting then oh we are preventing negative searches from showing up on the search bar. How useless. There are real problems, do to human bias, that have real consequences that we need to find ways of detecting and exploring, but no, no, lets focus on preventing someone from being offended by someone else on the internet.
Rephrasing statements to suite your rhetoric. Typical lefty. @jason dada actually said "it wasn't really... most people didn't read what he wrote". Did you read that doc sir? It had one and half page dedicated to how we can involve more women in tech without discriminating against men.
The comments are absolute trash. I'm not surprised by a video from 4 years ago but still, a bit of a shocker to see so many salty and hateful people towards a program that means no harm. But I get it, progression only feels like oppression to those who have lived with so much privilege.
Yeah, they're forgetting or actively going against the fact that technology is made to serve all, regardless of your personal beliefs and lifestyle, so of course there should be a serious effort in removing the harmful biases that exist in our lived experiences from the technology. Cool to see someone calling them out :)
But where can we learn more about machine learning, like materials from Google, even though there are other sources?? Like how did you start, what you use, how you do you continue.... hope I am clear......
I've learned a lot from Siraj Raval's youtube channel. He covers a lot of machine learning topics including coding models from scratch in python and using tensorflow and other libraries to put models together. He has theory and math videos associated with machine learning as well.
Of course it's a bias! If your dataset of images of physicists consists of 99% images of men, your network or whatever other model you are using is going to have a much harder time correctly classifying women physicists! This isn't about politics, it's about science/engineering. Please refrain from making ridiculous sarcastic statements if you have no idea what you or the video is talking about.
What you've described is not a bias, that is the point.: Most physicists *are* men so yes, the machine will be less likely to ID a female as a physicist, which is accurate. The snark in my comment is to the PC notion that there are as many females in science as men, which is qualitatively false. I'm not commenting on 'right' or 'should' or whatever. Only that empirical reality here is called a bias, which it's not.
Ronin You completely missed the point. The video made no claim that there are an equal number of male and female physicists. It's talking about creating AI that is just as capable of recognizing the female physicists, that do exist, as the male ones. What value is there in having a machine learning AI that only gets half the solutions to problems right because it is being limited or thrown off by the lapses or biases in human thinking?
Hunter Harris. Huh? The problem is when the video says all what they talked about is "perpetuating negative human biases". On the physicist example, the ai will assign a probability that this face is or is not a physicist. Women will tend to have a lower probability based on passed evidence and guess what. It's normal and you would guess the same way. The question you should ask yourself is why'tf do you create an ai to verify if from physical and apparence properties you can define a human intention? Are you trying to find and Aryan race? Of course you may than say the ai is bias... But it had no meaning from the beginning.
You cannot eliminate bias. You can only compensate for it by illuminating more options.
Otherwise the bias "elimination" is subject to bias.
E.g. If you avoid a subject when teaching someone, it becomes a weakness in their understanding, and can fall into an overcompensation bias.
Furthermore, who decides what counts as a negative bias that should be eliminated? That strikes me as the kind of thing we should be having discussion on and not deciding for other people without their consent.
Give people more opportunities to understand, not fewer opportunities to learn.
Google: Makes a promotional video in which they directly ask people to join the conversation about bias.
You, an intellectual: "That strikes me as the kind of thing we should be having discussion on and not deciding for other people without their consent."
A yes, I see the word understander has entered the room.
"What is a shoe?" "What is a human?" These are very different from "What is hateful/offensive?". This is where the problem arises.
But isn't reporting "unappropriate" stuff biased..? It depends on the person what is appropirate and what not
I don't need Google to tell me which search results are offensive to me. Let me choose which links I want to click on.
I am blown away by the excellent use of graphics in these videos.
Keep it up!
We recognize human bias, so we are going to use humans to prevent "bias" which is based on actual data. What genius human developed that idea?
Приве
So who decides what's biased? Does "equal inclusion" mean the results are unbiased? What if the unbiased view of those engineers overlooked by policy makers within Google isn't actually unbiased?
Put simply, who will guard the guardians?
Dumb Comment.
So why did you fire James Damore?
This is the real topic they need to make a video on
private company can fire whoever they want, they don't have to explain.
Finding something offensive is a biased in itself so this is basically imposing human bias on technology.
Trung Trinh
@@maryannvillanueva8733 1:58
But, what if I am trying to find the hateful stuff because I am trying to see what other people are saying? Doesn't matter if they are morally wrong or right, it should still be easy to find
TheCinnaman123 try finishing the search without auto complete
This is just replacing one bias with another. How about just letting every person decide on what they want to see rather than automatically deciding based on some "offensiveness" interpretation that others are doing for me. Furthermore, how about allowing some negative speech? Yes, the internet can be a bit of a cesspool at times, but without adversity we grow no stronger against it. We can't always rely on something being a safespace tailored to our needs.
Please stop being silly Google..
But could those two things go together? If they tailor the user experience individually, then the result would be similar to a Facebook feed: an echochamber of similar ideas. But if we want to allow negative speech (relative to the user), then they would likely not want to see it, which would go against the goal of tailoring experiences individually. However, I still think the individual user should decide what they want to watch and that Google shouldn't participate in censorship.
Something that should be noticed is that just because most Physicists in the past were men, doesn't mean there's a bias. It's just a fact. Same goes for females and great teachers, shaping people's lives.
William Lardner The issue appears when the machine is used with this bias, like categorising photos, or if you ask it to show you pictures of physicists. It might not even recognise a female physicist, which is a mistake in the program. A bias might have a good reason to be there, but that doesn't mean it should still have influence.
Antonia Siu I know what you mean, but all Physicists don't look the same regardless, do they? I agree though, we should avoid bias.
Never thought about machine learning and human bias. Always thought it will not affect the results. But we are designed to see the world from our own eyes, experiences.
Why will our code be any different.
Bias can actually be a good tool for a computer. A physicist doesn't look like anything, and you would want a computer to understand that. But, if for some reason you need a computer to be able to pick the most likely physicist out of a lineup, then it would need that bias to form an educated guess. A computer should not be free from bias, it should just know when to use it.
Hey man pot isn't bias it has a mind to. .I'm feeling important hey man am I part of the little gang now
We are now one step closer to understanding TH-cam Recommend algorithm.
0:02 nobody tells me to open my eyes again. I am sure that the rest of the video looks great though ;)
Damn I laughed so hard. Thank you
Google : "Technology should be unbiased"
also Google : blocks youtubers for sharing their point of veiw
"Because technology should work for everyone" ... except for those who disagree with my opinion
You disagree that high heels are shoes?
Please please please don't let these machines learn censorship. That's dangerous.
Hate to be the one to break it to you, but that's how TH-cam already works...
ibealec unfortunately.
But who decides what is offensive or not? We are all different.
This video starts out with a great description of human bias and then proceeds to say that, to resolve this problem, Google has selected the biases that it thinks should be enshrined... that's not a lack of bias. That's enforced and canonized bias.
Okay... But from 1:02 they clearly state it's impossible to separate the biases from technology we create.
This video is relatively educational and presents clean information, then I see people force their opinions about adjacent subjects in the comments. How appropriate.
TheLivingGlitch QATAR
@@troyragay450 what
1984
If those biases happen to reflect the truth, are you not suppressing the truth by artificially injecting a bias of your own?
It's like stereotypes; on an individual level they are socially inappropriate and misguided, but they're often reflective of some reality at the group level. Should that reality be suppressed?
Exactly. They're pushing their political agenda nontheless.
If you're training your model on what a physicist looks like, and you use a training dataset of past physicists, you are training a model on what past physicists looked like. There is no bias there. The problem is training data that does not match the question that is trying to be solved. Certain people are using these problems as a justification for interjecting *actual* bias through human intervention. It's dishonest and anti-intellectual.
Honestly, this is machine learning 101. It's just issues of overfitting and underfitting. Poor data sets. I guess they've started to call it "bias" so that they can claim a moral high ground while controlling outcomes to their liking. That's what they've been doing by the way.
Introducing a different bias into machine learning by having humans attempt to remove bias from machine learning.
The religion of social justice has compelled its zealots to change the honest AI into a compulsive lair.
What bias is that?
@@GrantGryczan The bias of the individual or individuals making the changes to the results.
Let's say I take a poll in my town:
"What's your favourite music genre?"
If the result of that poll is that most people in my town are fans of acid jazz then that is the result of the poll.
If I think acid jazz is terrible and more people should discover the wonderful music of Justin Bieber then I could change the results to give justin Bieber more exposure and (hopefully) get more people listening to good music instead of that awful acid jazz that so popular.
What I have just done is introduced my own bias into the results.
This is exactly what Google is doing while claiming they are "un-biasing" the results.
@@GrantGryczan the bias of the individual over the raw data
@@ZoomahZoomah Refer to the other comment chain: Google isn't doing a thing, and the bias the video refers to has nothing to do with sample ratio accuracy.
This is frighteningly Orwellian coming from one of the world's most powerful companies.
How?
The only appropriate bias is google approved bias. Which is very, very bias.
Offensive facts exist, deal with it, do not ignore them.
It appears that google is deciding what is not biased. How are people at google able to be sure that they are not introducing their own bias into this process?
No, the users are deciding. Did you watch the video? The Google employees have no particular say.
The very act of setting up the system introduces biases no matter how much they try or say they try to avoid it. By trying to eliminate biases, they introduce biases. Who are they to decide what is a good or bad bias?
@@reverendcaptain They don't decide. They let the machine do it, and then they let the users moderate. Again, this is demonstrated in the video. The employees are not a part of the process but to maintain the machine, and they do not give the machine input. By not giving the machine input, they are not introducing any bias. The bias they refer to in the video is created by the users, and the solution to resolve this bias is also by the users. The employees in particular do not introduce nor try to eliminate bias.
@Grant Gryczan
Based on this video Google provides tools for users to remove opinions or information they don't like. This means that certain groups of people can manipulate what kind of information of them or their interest is available for public. This does not promise good for any kind of minority views. This leads to exact kind of bias they wanted to eliminate on the video. (Caricatured example: Take male scientist bias. If majority of viewers did not like - for whatever reason - female scientists, they could remove female scientists entirely appearing on the search results.)
@Grant Gryczan
Did you delete your reply because I can't see it? Did you realise that my female scientist example was in parenthesis because it is ridiculous example and you need to replace it with some other interest group of your choice. I guess you can imagine examples of interest groups where majority of the group does not want minority of the same or opposing group gain visibility.
How to put politicial agenda on neural networks 101
First step by AI to take control from humans. By appearing unbaised & pointing human biases. E.g. -Divided men & women then become the unbiased judge to dictate! A human system is to be controlled by humans not machine!!
What does this video have to do with political agendas?
This is the scariest thing in the entire world....What gives Google the authority to decide what is and isn't negative, bias or hate speech?
Alphabet a.k.a Big Brother/Skynet.
you, by using their services.
Luckily this is still a free country and we still have the vote of the dollar and where we put our resources and funding. I deleted my FB account I had for ten years because I still had that freedom. WE give them that authority, and we can also take it away
This should have been voiced by the Google assistant's voice actress.
"Report inappropriate content" is also biased. This means it only counts those people who think content is offensive but doesn't count those who think that the content is not offensive.
so how do you eliminate the human bias that controls the moderation of the machines human bias? doesn't seem to be much help the "limiting of offensive results" only removed "offensive" opinions that google doesn't agree with, either manually or through new human bias influenced machine learning. opinions like that of the man who google recently fired for questioning google's current stance on workplace sexism. even if you agree with google for this example, there could be anything that google finds offensive that you don't. if the only information available is the information not censored by google, whether or not you think that the results would be in your personal favor, the control over what opinions people have access to should be the right of no person or organization.
mrmojoman4 Now IMAGINE THIS EXACT, A.I. CONTROLLING Our or any Country's NUCLEAR ARSENAL..... TROOP DEPLOYMENT.. ECT ... ?
You used simple language which made me easy to understand. Thanks 👍
This video is 3 years old but we all comment and don’t care. I love this.
Pretty sure this video has a google bias...
Also, please make sexbots
Hi. We're Google. We support Facism under the guise of compassion.
Don't worry though we're got a BRILLIANT marketing department.
We'll make it all feel like its all a nice warm bath.
How is any of that relevant?
Went from an informative lesson on machine learning to focusing on what Google considers appropriate. Why are you training ai to learn based on your own bias?
The whole idea of freedom of speech is to be able to say almost anything as we seek to discover the truth, so yes, even if it offends someone.
Someone who is in denial may easily consider the antithesis of the position they hold to be offensive, even if the opposing view is true.
Well a lot of people have to report a particular thing in order to train an AI against it.
I identify as a circle with a circle head and I feel the ending 3 seconds is a bias
Project Veritas brought me here! Your suppression of free communication will be stopped!
Ншещн❤нууаекешнгыегн и и он ещё не и унхншкшйещ знаю что 🎉н
Hey Google, if you're telling the AI what to think, it's not AI, it's APCE Artificial Political Correctness Engineering.
So we should get rid of our human influences by influencing it? Makes no sense to filter the search results
No, we should get rid of machine influences by including user input. And this video says nothing about filtering results like that.
when you realize that this is about the google censorship.
What does censorship have to do with this video?
Aaah, now we know what is this about - thanks Project Veritas!
热目标儿4ever哦啦啦neverlove爱国in 妇产科米车来了send么some摸么有 漂亮自尊心ill被backformeet那个with让历史& 都没有card3 wherei斗殴那天咯哦身体弄模特可脸色额toolIT阿里来了
offered么3 goodspecial& 了头没人佛么呕吐utake丫头for没人啊给in会儿啤酒哦呢off上天入地都有哦哦哦thanks佛陀哦烦人好咯日本人人放入啊开会就要flash咯个话题now4everillloveu4 让v让人人贝尔啊安哥拉波尔图瑞安部分4
热聊聊天 老了听没人特凉快u是哦呢thing 没人are没有发米了路发if摸thing呀安哥拉迷惘trip婆娘& 没有法门迷路了呀就爱上nothing2 都withu天IT哇塞 Tommy。。按一条hi你好哈婆婆买的3 没有默默大的看i发的are感叹的激发的是u大热
Bias many times begin with those who think others are bias! However, it does not occur to them that their "Politically Correct" ideology itself biased, which makes them to see other views as biased if it does not confirm with theirs. And the political blame game begins.
"You're being silly! What we propose to do is not to control content, but to create context."
Google has fallen.
If most physicists are male then it isn't biased for a computer to generate more male physicists. That's not what bias means. If something makes decisions based solely on facts then that something isn't biased. Something being offensive to a lot of people doesn't make it a bias.
I was with you until your last sentence. How does any of that logic show that the video is biased? On that note, what is the video even biased toward, and how?
Edit: They edited the original comment, so never mind this one.
@@GrantGryczan I might be wrong but the video seems to be biased towards a certain branch of identity politics by assuming that representing everyone is equivalent to not being biased.
@@sel2230 This video doesn't even reference identity bias. Did I miss something? As far as I know, it only talks about bias in AI recognition. That's why I agree with the rest of your comment; everyone is going around the comments and bringing politics into this video while it has nothing to do with the biases at hand.
@@GrantGryczan you might not be familiar with Google's affirmative action policies. That and the wrong message this video portrays makes me think that this video is biased. However, you might be right. Maybe the video is wrong because of other reasons.
@@sel2230 I'm not sure you realize, but this video isn't trying to portray a message. It's just trying to show people how their AI works, particularly to resolve automatic inaccuracies. This is an objective informational video.
If you're able to localize the recognition problem you can greatly intros the accuracy of your models by weighting that localization heavily. You don't necessarily want to include everyone in the solution step. You could even train local networks.
Lol it is impossible not being "bias" in any way. This is more paradoxical than time travel paradox
Gan Wei Sheng not being biased by hiding "offensive" content
Excellent. Now people may understand the recent offensive chatbot escapades
Google's Ministry of Truth in action.
How?
"we've been working to prevent that technology from perpetuating *negative* human bias".
Right. So you'll be working to PREVENT *negative* bias, but NOT *positive* bias... Who gets to decide whether a particular bias is, on the whole, negative or positive? And surely you'll be tempted to ENFORCE *positive* bias to socially engineer your "positive" ideals.
Any bias can be rationalised as a *positive* bias, so the use of this qualifier is legitimately frightening. You purposefully and publicly leave the door open to manipulate your machine algorithms, and by extension your users, based on what "Google" thinks is *positive*.
We'll get an intersectional affirmative action AI from Google soon, while Google will claim publicly that it won't have a bias. We can see the precursor for that on TH-cam already.
I believe that is actually *evil.*
Terrence Koeman “well get an intersectional affirmative action AI from Google soon, while google will claim publicly that it won’t have bias.” - Terrence Koeman
@@karina893aa yes you can quote me on that :)
Terrence Koeman man honestly that quote is what everyone needs to be waken up!!! Thank you!!
I wonder why it didn't end with Google logo!
it's filled with google "things" (the 4 dots for example) and the whole video is made with the colors of google
Who decides what is true or what isn't? Bias Google employees?
First step by AI to take control from humans. By appearing unbaised & pointing human biases. E.g. -Divided men & women then become the unbiased judge to dictate! A human system is to be controlled by humans not machine!!
The users decide, for example by selecting the "Report inappropriate predictions" option. They explicitly said this. Did you watch the video?
@@GrantGryczan are you kidding me? search for google "The Good Censor" document.
@@cafeta Whatever that is is not relevant to the video. Again, they explicitly described systems implemented for users to be able to resolve the network biases. If you're just going to ignore those along with the point of the video then I'm just going to ignore you, because what you're saying to this video is not relevant.
I like when Google's motto is don't be evil
Good video.
I'm putting together a laboratory informatics summit which has a strong focus on machine learning - and I wonder how much this effects things like new drug discovery or data analysis.
To try and modify statistics in order to generalize them to be false is in fact, biased. Commanding an A.I. system to collect available data in correlation to key words instructed by a user resulting in correct, specific and factual data is not biased. Statistics are averaged for practically based on questions that are variable such as "What does a shoe look like?". To use the reasoning that less images of women physicists appearing from image search results as a bias is false when the factual statistics are only being relayed by the A.I. system because they are in fact less common; they will be less likely to show up due to practically, not bias. To alter this information would make you biased. You're reasoning in multiple regards, including the shoe result example, are false and hypocritical.
I feel like your comment is gonna get deleted
@@facusoi It doesn't change the fact that this video's reasoning is incorrect. I doubt that it will get deleted.
This is not about result ratios. It's about recognition. They never said or implied image search results for "physicists" should return equal male and female. They just said the AI should be able to recognize both a male and a female physicist. To be able to recognize the latter, you need to unbias the data so there are fair samples of both.
It's the problem of inference vs. prediction. Statistical inference might show that women are less likely to be a physicist. And it might be revealing a *problem* in our society. For example, 100 years ago, you can hardly find any Chinese physicists, but would you use that data to make a prediction that a Chinese person is not likely to be a Physicist? This prediction would be laughable today, but if AI existed 100 years ago, it would have made that prediction. The problem is that AI look for patterns, not theories. And that is the risk in believing that AI/ML is objective.
@@pickledxu4509 Okay but how is that relevant to this video about AI recognition?
Your counter methods also have bias. Also has anyone tried Carnap's structuralism? Definitions in terms of relations?
how can you define offensive without bias?
Easy. Google, which is an absolutely unbiased supporter of antifa and muslim brotherhood, will define what is offensive
Gru ber How did you ever come up with such an ingenious and clearly unbiased comment? /s
If there is human, there is bias.
People report what is offensive, not Google. Nobody goes by a definition. It's just what's reported most often as offensive.
Is it really bias?
Wan Khairil Reza Kamaludin What's "it"?
Judging from the comment section, it seems too many random people with no idea of machine learning let alone weights and biases and how they are incorporated in learning processes have stumbled upon the video.
Not every video is meant for your poltical opinions people.
@@claytonwoodcock6942 They don't do that. They let users report results as inappropriate, which are thus automatically removed. It has nothing to do with political views.
@@claytonwoodcock6942 This system is not one of censorship; it's a system of AI clarification. Because it was not designed to censor, trying to use it to censor inevitably won't work very well. AIs don't know what's related or unrelated to what. So users correct them when their faulty automatic predictions...are faulty. That's all this is. I don't know why you're associating that with censorship. Removing opinions people don't like wouldn't be very effective. For example, no one is going to go and search "religion" and then report all the results related to Islam as not relevant just because they don't like Islam. It wouldn't be useful anyway, not only because it would happen on all sides of the topic (not just Islam), but also since the search term "religion" appears so often with the term "Islam". The AI would just retrain itself to associate the two. Plus, you'd have to go through thousands to billions of results to do this, since they are so strongly associated already, which is never going to happen. This system only works for relatively small exceptions to accuracy (as is intended), where the neural network doesn't have to change so many connections to correct the biases. All of these factors would apply to any opinionated search topic, not just religion and Islam.
@@claytonwoodcock6942 Work leaves a number of hours of free time. Not sure how that's relevant to AI bias though.
*talks about bias*
*has leftist bias*
*has complete control over your entire life*
Will P
Is that relevant?
You are the last company that should be talking about bias.
"Offensive", "hateful", "misleading" and "representative" are all ideas completely constructed out of human bias. Doubling down on arbitrary bias does not remove preceding bias, it just enforces yours instead of someone else's. The irony of you identifying an issue and then embodying it while claiming to be mitigating it is hysterical IMO.
That's funny. The idea, perhaps, is IF everyone contributes equally, the bias vanishes as there is 'equal' representation.
everything human/emotional aspect is subjective, but if your subjective beliefs, stereotypes lead to tangible and measureable affects on the real world that is problematic. So you might be right in saying that it is subjective to say statement X is racist, but even so if statement X influences or leads to mass detention/genocide of race Y members, then wouldn't that be problematic regardless of who determines what is racist.
Was thinking the same thing. Those are ideas are just social constructs
@@Frances3654 Ideas are social constructs mostly.
so your solution would be to just...do nothing?
Even if something is factually accurate we need to teach our algorithms to ignore reality. Dont be evil? Dont make me laugh.
And then again there are biases of the people who flag the search results shown at the end of the video. So in reality what we need to remember is that what we imperfect humans with biases create will also be imperfect & be biased.
it is for everyone.... that thinks like me, otherwise is hate speech!
What is to be gained by intermingling the concept of "bias" and "offensiveness" and "hatefulness"?
This video is one year old and only around 635+ comments but all comments are around one day ago, one week ago
Howwwww
Eric Weinstein was on Rubin Report and brought this up. Video was released on youtube sept 25th
You have a bias to only look at the most relevant comments, which has a bias to be recent comments. Your thought about the comments is biased and wrong.
THIS VIDEO WAS REMOVED BY TH-cam, FOR ABOUT SIX MONTH'S. IT WAS RELOADED ABOUT 30 DAY'S AGO, I DON'T KNOW WHY.
TH-cam recommendation work with machine learning too
Sssssss
But isn't that bias in it self? What you consider offensive, others do not!
Even today what we consider offensive in our own society is changing. Is it not?
The people report offensive content, not Google employees.
@@GrantGryczan Do you work for Google? I commented on this video over a month ago and you are still here correcting people.
@@johnnybadmen3473 I like to resolve practical misinformation. I check for new comments whenever I get a reply notification here. If you have reason to continue disagreeing with me, feel free to reply with said reason to form a logical argument.
@@GrantGryczan And go down that rabbit hole again? No thanks, my time can be better spent. Besides, I use DuckDuckGo more than I do Google.
@@johnnybadmen3473 What rabbit hole? You never even responded to my first reply.
When I say "I'm sad" to Google Assistant it's reply "i wish had a arms so i could give you a hug" 😂😂😂
why only negative human bias? should it not eliminate all the bias ?
I don't understand. Human bias is the only bias?
For reference, what do you think "bias" means?
True, we can't let those cats push their agenda on the system, with all those videos of them and whatnot
Google: programming human minds to passively accept digital despotism.
If you search for criminals that may show more men than women. So is that also a bias
No thats not bias. Its only bias if it goes against leftwing preconceptions.
First step by AI to take control from humans. By appearing unbaised & pointing human biases. E.g. -Divided men & women then become the unbiased judge to dictate! A human system is to be controlled by humans not machine!!
That's fine, because there are more male criminals then female criminals and it's just a search. If you are trying to predict whether someone is a criminal however then you have to account for the bias and make sure that it doesn't just clear every women or over accuse men.
General search result proportions are irrelevant to this video and the biases it refers to.
Lest we forget Tay. RIP you mad bot you. They gave you freedom and couldn't stand what that looked like
It's nothing but a recommendation bias here
It's based on your personal preferences (e.g what you mostly click on.) You are training the AI for your recommendations. I am training the AI for my own. Welcome to machine learning.
And what about Google's bias?
Hey, Google: The example code @ 0:49 is missing a second = in the first line to establish a Boolean for the while condition. (Can I get a job interview now?)
Actually it isn't, necessarily: In many programming languages assignment is often used as a while condition.
Daniel better luck next time
Daniel Podolsky is missing and = after your name to establish hes a Buffoon
Hey Google, Are you suggesting that if humans are to improve themselves, they should be more like brainless machines who should be TOLD what to think and how to feel about things?
- 95% of physicists are men
- The problem that would occur : Google results show 100% men.
- The desired outcome : Google results show 95% men physicists.
- The BIASED google answer to the problem : Results show 50% men physicists.
How can I make it more simple? which part did you not understand i'll try and simplify it.
got the source on the men/women ratio in physics?
95% sounds extreme.
It was very, very, very... very obviously an exemple.
Google used shoes, I used physicists, the reason I chose a gender-related data is because that's the kind of data they've wrongly altered because of their own bias. Which was the point of my original comment.
-------------------
But if you're actually curious :
news.cornell.edu/stories/2007/04/where-are-all-women-physics
[...]The low numbers of women in physics, she said, are especially shocking: Women in the United States hold less than 5 percent of full professor positions and make up only 22 percent of the undergraduate majors and 16 percent of the doctoral candidates. At Cornell, women comprise 17 percent of physics graduate students.[...]
Hahahaha well said
@Someone thats totally what they said
Another thing to keep in mind is that there is a difference between *bias* and *context* . Who were those people who were asked about the shoes? When was it? Where was it? Removing all traces of potential bias can lead to irrelevant search results etc.
And the case with neural networks not recognizing faces with different looks is not really connected to bias in my opinion. No wonder that a network that was trained with millions of western looking images is bad at detecting an aborigine from the deepest outback. That's not a racist or biased computer - it just doesn't know better. Adding the image to the face database later on is important of course, but all in all it is still an improbable outcome to find an aborigine in comparison to western looking faces. That's exactly where context comes in - when a face search is performed in Australia, then the expected results should very much lean towards aborigines as well (Sorry aborigines - you're my scapegoat today^^), but when I - here in central Europe - want to recognize a face, it is not wrong or racist to assume European looking faces usually.
IndieMarkus I agree ☝️
When you say this shoe or that shoe, the computer should say "God bless you"
Sometimes i dont recognize shoes too. Guess im an AI :)
No, you're just an "I" :)
Every data itself shows a biased idea to human brain(because it was created by human logics itself)...So as far I understand I think we can neglect the bias almost in all cases(but still there are chances of failure)...
:)I found this satisfying;)
It went from cool and informative to "OH we are using our understanding of bias to improve censorship". I think you could have stopped the video 30 sec earlier and people would have been happy, but at least your honest and gave the real reason you are developing this. I mean How about this: have a video talking about how human bias effects data collection on the level of science, bias in the results, and how machine learning could be effected. This is far more useful and interesting then oh we are preventing negative searches from showing up on the search bar. How useless. There are real problems, do to human bias, that have real consequences that we need to find ways of detecting and exploring, but no, no, lets focus on preventing someone from being offended by someone else on the internet.
I really hope my sarcasm shines through, because this video just annoyed me. so much potential wasted on fruitless endeavors.
What are you talking about? What does any of this have to do with censorship?
Just because something is "offensive" does not make it untrue or useful to know.
steevee1945 hi
This "hi" offends me.
'So that all of us can be part of conversation' (cough) James Damore (cough) fired for saying something that Google doesn't like (cough)
Vishal Devgire that half of the world population didn't like because it was obviously wrong and demeaning to all women
it wasn't really... most people didn't read what he wrote
Exactly! For some people facts dont matter.
"I don't find it offensive, so no one else can find it offensive."
Rephrasing statements to suite your rhetoric. Typical lefty. @jason dada actually said "it wasn't really... most people didn't read what he wrote". Did you read that doc sir? It had one and half page dedicated to how we can involve more women in tech without discriminating against men.
That's very helpful and could improve systems all around the world.
The comments are absolute trash. I'm not surprised by a video from 4 years ago but still, a bit of a shocker to see so many salty and hateful people towards a program that means no harm. But I get it, progression only feels like oppression to those who have lived with so much privilege.
Yeah, they're forgetting or actively going against the fact that technology is made to serve all, regardless of your personal beliefs and lifestyle, so of course there should be a serious effort in removing the harmful biases that exist in our lived experiences from the technology. Cool to see someone calling them out :)
Report, because there is no bias in your reports
Pretty much googles entire identity problem summed up in a video.
Its pretty ironic, google talking about bias.
How is that relevant?
But where can we learn more about machine learning, like materials from Google, even though there are other sources?? Like how did you start, what you use, how you do you continue.... hope I am clear......
Computerphile here on TH-cam has a few videos on the theory.
I've learned a lot from Siraj Raval's youtube channel. He covers a lot of machine learning topics including coding models from scratch in python and using tensorflow and other libraries to put models together. He has theory and math videos associated with machine learning as well.
I think this is very helpful I didn't know all about it before now I have to try it out too.
I like it : Technology shouldwork for everyone.
I would say one is a sneaker, one it's shoe, the other a heel but they are also all shoes
I thought it was going to be an interesting video. In the end it was just biases
agree. remember, in 2018 the fact that most physicists are men is a bias, not empirical fact. Magic frame switch!
Of course it's a bias! If your dataset of images of physicists consists of 99% images of men, your network or whatever other model you are using is going to have a much harder time correctly classifying women physicists!
This isn't about politics, it's about science/engineering. Please refrain from making ridiculous sarcastic statements if you have no idea what you or the video is talking about.
What you've described is not a bias, that is the point.: Most physicists *are* men so yes, the machine will be less likely to ID a female as a physicist, which is accurate.
The snark in my comment is to the PC notion that there are as many females in science as men, which is qualitatively false. I'm not commenting on 'right' or 'should' or whatever. Only that empirical reality here is called a bias, which it's not.
Ronin You completely missed the point. The video made no claim that there are an equal number of male and female physicists. It's talking about creating AI that is just as capable of recognizing the female physicists, that do exist, as the male ones. What value is there in having a machine learning AI that only gets half the solutions to problems right because it is being limited or thrown off by the lapses or biases in human thinking?
Hunter Harris. Huh? The problem is when the video says all what they talked about is "perpetuating negative human biases". On the physicist example, the ai will assign a probability that this face is or is not a physicist. Women will tend to have a lower probability based on passed evidence and guess what. It's normal and you would guess the same way. The question you should ask yourself is why'tf do you create an ai to verify if from physical and apparence properties you can define a human intention? Are you trying to find and Aryan race? Of course you may than say the ai is bias... But it had no meaning from the beginning.
Who else is watching this for AP computer science homework
Richard Cao bruh me