ITT: We design machine learning for youtube that doesn't suck

Is it possible? Seems like their demonitization bot doesn't even do feature extraction.

Why would you say it is like kicking dead whales down the beach? I'm sure it's making them plenty of money and that's pretty much all they care about. Adjusting the heuristic/metric to optimize would do a lot more good than giving it more features to work with.

>>3879 Not sure if "kicking dead whales down the beach" is a replacement, or if I'm just not getting your response. Google has implied that however the algo works, it has no understanding of *why* a movie is inapropriate, and therefore can't give feedback to users. Appropriateness, apparently, is the only metric it is given. It'd be 100% less evil, to teach it to detect "gore" versus "collection of pixels statistically similar to a corpus of removed videos who share nothing in common besides their removal"

>>3884 It's a text enhancement.

>>3867 what kind of information do you have in mind for extracting with ML?

>>3891 For reference, we are talking about Google's practice of using ML to flag "inappropriate" videos. They have basically gone out and said the algorithm they use isn't capable of categorizing the offending content. For that reason the content creator doesn't ever get to find out what he did wrong, and the false positive rate is immense. So I would want to extract out the type of offending content. Distinguish between violence and "controversial political" content and whatnot. The current implementation is so bad it's flagging on blank videos, and videos of dogs and Google claims that even *they* are not made aware as to why the false positives are happening, besides the algorithm requiring more training data. If the algorithm was able to extract features, like the *type* of offensive content, then they would be able to offer constructive criticism to their content creators.

I think it comes down to the fact that you can't reliably do the flagging of offensive content via ML. For example, how would you differentiate real violence, like the recording of a shooting, from those action sketches (which sometimes have blood in them)?

>>3901 ML is *supposed* to be the magic thing that let's you do that kind of thing, because it answers the "how" for you. Some might say that if you already know how to find the answer ML isn't always the best solution. Anyway, at least in my thinking getting FP'ed for "violence" rather than "undisclosed similarity in the linear algebra" sounds a lot better.

>>3902 t. Person that has never used ML themselves.(Elaborate; avoid spilling soykaf.)

>>3903 Constructive(Elaborate; avoid spilling soykaf.)

>>3902 >ML is *supposed* to be the magic thing that let's you do that kind of thing, because it answers the "how" for you. The only difference between a movie scene where they simulate murder and an actual murder is fake blood. Even a human cant tell the difference as movies can be very realistic. There is nothing *magic* about ML, it uses the same deductive reasoning a human does. No offence, you dont know anything about ML, so you have to start at the beginning and read beginner ML books. Before you can start on ML projects you have to learn exactly what ML can do before you even think how to apply it.

>>3906 I hear a lot of salt and zero technical knowledge, and it sounds like you are trolling to get me to argue with you. I'll take the b8. Maybe the term magic triggered your autism? Bro, if you already know how to solve a problem then you don't need ML. E.G "If I already know the formula for what my customers like so I don't need an adaptive recommendation system" or "If I know how to fully optimize this process I wouldn't be training this neural network to do it better than I can" If you knew exactly how to do a thing you wouldn't be using a damn heuristic, so stop being a cunt. And you aint said soykaf to refute the basic idea their soykaf doesn't do feature extraction or suggested a design for a better way. You know about feature extraction right? https://en.wikipedia.org/wiki/Feature_extraction That's a real technical term. I used one instead of just being a flame spewing troll. In fact dude, I could be a complete idiot with regards to ML and you would still be able to fuarrrking SUGGEST a solution as the original thread suggests. I'm sorry whatever bank in Indiana you work at doesn't recognize your genius, but you could be a little gracious and respond to the thread instead of trying to talk smack.(Always argue in good faith and avoid using personal attacks.)

>>3920 >Bro, if you already know how to solve a problem then you don't need ML. That would imply there are problems that computers can solve but humans can't. I don't think so. Check your logic before exploding into childish rage.

>>3923 Bro post some tech details before exploding into childish smack talk. I just see trolling here.

>>3920 >>3933 Im not trying to shoot down your desire to want to do ML on youtube videos, but the specific example you provided is almost impossible to do. There are people who work for the CIA/FBI/NSA who's entire job is to examine photos/videos to figure out if an actual murder was committed or if it was staged. Its almost impossible to find any kind of visible patterns in a photo/video to figure that out, so using ML is not going to help. Yes there are plenty of other areas where ML is very easy and practical to use, like in what you suggested above: >"If I already know the formula for what my customers like so I don't need an adaptive recommendation system" Thats very easy to submit purchasing data and come up with patterns for different types of customers. But when you are talking about analyzing photos/videos, that requires an entirely different level of sophistication which is way beyond just feeding data to an ML algorithm

>>3938 Bro, I don't think you read the original post. I don't want to do any ML, I want to collaborate to propose a better way for youtube to do their *currently existing* ML. >There are people who work for the CIA/FBI/NSA who's entire job is to examine photos/videos to figure out if an actual murder was committed or if it was staged. Its almost impossible to find any kind of visible patterns in a photo/video to figure that out, so using ML is not going to help. That is so not a relevant or conclusive anecdote. Bro, you can so extract features like "smile", "bird", "blood" etc... from video. "Man wearing ku klux klan" hood shouldn't be too hard either. Check out the entire journal: "Recent Advances in Visual Information Systems" Here's a fun looking paper: http://ieeexplore.ieee.org/document/6909728/ shows just one example. I mean dude, I've fuarrrking seen people whip up the style of algorithm I'm describing FOR VIDEO in person. Stahp with the uniformed BS. If you wanna poo poo the idea than give me some science like: "non negative matrix factorization based machine learning works well at the small scale, but could not possibly work at an acceptable cost point for Google" etc... The idea is to embarrass them by showing how trivially they could cease to be assholes alright? Jeez, the way you are shooting down the entire idea makes me think you are a student, or you work for Google.

>>3933 >Bro post some tech details before exploding into childish smack talk. I just see trolling here. Your logic implies that there are problems that ML can solve but humans can't. All you need to do is post a counter example. Since you can't think of one I guess >>3923 was correct and you are wrong.

>>3963 >Your logic implies that there are problems that ML can solve but humans can't. All you need to do is post a counter example. Since you can't think of one I guess >>3923 was correct and you are wrong. What are you a philosophy major? I just posted an example of the type of state machine you'd use to do the very task. Shut your dirty mouth. And I never said that there was a class of problems that machines could solve that humans can't. You Autist, what I said was that you use ML when you don't *know* the answer - as in you are too assed to find out on your own. Start talking state machines if you can. Otherwise go back to your liberal arts textbook.(Always argue in good faith and avoid using personal attacks.)

>>3973 The reason I'm posting is because you seem to be frustrated that no one is taking your thread seriously. >Bro, if you already know how to solve a problem then you don't need ML. >I never said that there was a class of problems that machines could solve that humans can't Saying dumb stuff like this is part of the reason why. >Start talking state machines if you can. Otherwise go back to your liberal arts textbook. I don't think you should dismiss the advice in >>3906 to actually learn a thing or too about the subject before starting a thread like this. >I just posted an example of the type of state machine you'd use to do the very task. I don't see any state machines here. You posted a link to the Feature Extraction Wikipedia page and a link to a paper titled Efficient Feature Extraction, Encoding, and Classification.

>>3991 >Saying dumb stuff like this is part of the reason why. More baseless aggression. >I don't think you should dismiss the advice in I appreciate your devotion to deriding me, but in a thread where the topic is essentially "let's suggest a better way", saying "git gid dumbass" is more or less off topic. I promise, if you contribute in good faith, by actually talking about the subject I posted, I'll read a textbook. Feel better? >I don't see any state machines here. You posted a link to the Feature Extraction Now this is the part where you are 1) getting autistic and 2) are revealing that *you* don't know anything about ML. Motherfucka' I never said the link was to a state machine. I asked *you* to talk about state machines rather than just telling me to git gud. And if you don't know that feature extraction is part of machine learning activities when it comes analyzing media, then you know nothing of the topic yourself. The whole point of *linking* it was to demonstrate that to you so you'd stop deriding the whole idea as "impossible." Wanna know how Spotify's ML tells what genre an uploaded song is? Feature extraction. That's not a whole lot different than thinking about how to solve "is this guy a racist" But I haven't so much as heard a single tech term from you, or a real response to the thread that a liberal arts major couldn't have written. http://ieeexplore.ieee.org/abstract/document/6982196/ Here enjoy a paper describing a state machine. Did your parents hit you a lot or something? God.

Now, I've warned several of the posts in this thread. If someone is wrong, try to correct them without vapid insulting. It's on a different note that I find unease in considering how to improve a Google service.

>>4019 Fair enough. The only thing I disagree with about the warns you called out: I think just dismissing my entire thread in the way some of the posts do is equally non-constructive. I would argue that while I'm being ad-hominem, he's arguing "Argumentum ad lapidem" by dismissing the entire idea out of hand and not bringing any evidence to the table to support his ideas. I'm not saying I'm perfect, just that the aggressive dismissal is equally non-productive. Nobody has to post that kind of "sage" stuff, if we're all trying to have a nice discussion. >It's on a different note that I find unease in considering how to improve a Google service. Why? I'm genuinely curious. The point for me is not to help Google, but to demonstrate how embarrassingly easy it could be for them to add transparency to their demonitization-bot. Whatever you think of people making a living off youtube, I think we'd all agree we'd rather know the reason we were demonetized than not. And hey, if it's that guys opinion that it's literally impossible to design such an algorithm, I'd like to hear a *technical* discussion as to why. Technical on a level that would at least stand up to the standards of say Stack Overflow. I don't think his dismissals add up to a technical argument.

>>3867 Can you provide a resource on how the algorithm you want to improve works? >>3895 >So I would want to extract out the type of offending content. Distinguish between violence and "controversial political" content and whatnot. The easiest solution would be to train the algorithm to classify the videos. Instead of classifiying into good and evil it could classify into good, violent, controversial political etc. Off course, this would require you to provide enough training data for each class. If you want a further discussion, it would really be good if you could say how the algorithm works right now.

>>4015 >Motherfucka' I never said the link was to a state machine. >>3973 >I just posted an example of the type of state machine you'd use to do the very task. Shut your dirty mouth. There is a serious gap here between the things you think you are saying and what you are actually posting. You need to close that gap if you want a meaningful discussion. >The whole point of *linking* it was to demonstrate that to you so you'd stop deriding the whole idea as "impossible." More likely you Googled "ML + feature extraction + video" and posted the first hit. What >>3906 said still stands. You simply don't understand ML on a basic enough level to understand the objections raised in this thread. Instead you close your eyes and insist that ML is magic. >>3906 >The only difference between a movie scene where they simulate murder and an actual murder is fake blood. Even a human cant tell the difference as movies can be very realistic. There is nothing *magic* about ML, it uses the same deductive reasoning a human does. No offence, you dont know anything about ML, so you have to start at the beginning and read beginner ML books. Before you can start on ML projects you have to learn exactly what ML can do before you even think how to apply it. >But I haven't so much as heard a single tech term from you, Just dropping "feature extraction" and "state machine" into your posts over and over doesn't actually add anything to the discussion. >>4023 >Can you provide a resource on how the algorithm you want to improve works? I think that would be a good way to revive the thread.

>>4024 >Can you provide a resource on how the algorithm you want to improve works? All Google has ever released is that the algorithm does not classify the videos beyond "ad-friendly" and "not ad friendly" I don't think it's possible for us to criticize the exact technical details of their algorithm, and that wasn't the intention. I just want to demonstrate that they *could* do the right thing. >The easiest solution would be to train the algorithm to classify the videos. Instead of classifiying into good and evil it could classify into good, violent, controversial political etc. Off course, this would require you to provide enough training data for each class. Of course. I think Google has plenty of videos which could be trivially associated with the type of content that breaks their terms of service to form a training corpus. What kind of classification system would you recommend for this type of thing? https://static.googleusercontent.com/media/research.google.com/en//pubs/archive/42455.pdf It looks like Google research likes neural networks for the classification bit. Others on the net seem to agree. Here's another paper suggesting the classification step would start with a CNN: https://arxiv.org/pdf/1609.06782.pdf It seems pretty obvious to me that based on the stacks people are using, that training the classification step for very specific content should be possible. One thing I noticed is that most of the papers talk about classifying the video one frame at a time, which seems like it wouldn't work well/at all for movement based classification. Here's a guy who did his own interpretation: https://blog.coast.ai/five-video-classification-methods-implemented-in-keras-and-tensorflow-99cad29cc0b5 To quote his article: >Classifying one frame at a time with a ConvNet >Using a time-distributed ConvNet and passing the features to an RNN, in one network >Using a 3D convolutional network >Extracting features from each frame with a ConvNet >and passing the sequence to a separate RNN >Extracting features from each frame with a ConvNet >and passing the sequence to a separate MLP I have no idea of the method posed in this article is the most efficient method, but any way you slice it, it appears to be a very computationally expensive multi-step process. So far my conclusion has been that Google is either: a) Fully aware of the offending features of the videos b) Simplifying the algorithm to the utmost to save cycles. I admit my ignorance here, and so I am curious to hear what lain has to say about it. >>4024 I'm going to stop replying to you because you aren't productively participating.Your tone is on the verge of ad-hominem and you aren't backing up your claims with evidence. I will happily reply when you are on topic.

>>4033 >What kind of classification system would you recommend for this type of thing? The task sounds perfect for a neural network. As for the exact strategy, I have unfortunately never worked with video data. My first idea would be to interpret the video as a sequence of images. Then, I would take a convolutional network and make some of the layers recurrent. Convolutional LSTM units for example are a good way, to turn a normal conv unit into a recurrent unit. The first time I heard of conv lstms was in this paper: https://arxiv.org/abs/1511.06085 The main problem I can see with this is that the sequence would get way too long. This will make training almost impossible. I would solve that by only taking a small section of the video. If lets say 20 seconds of the video are flag-worthy, then the whole video is flag-worthy. Also, you dont need the whole context of a video to determine if one scene is bad or not. So I would always process subsequences of the whole sequence, where each subsequence has a maximum length (lets say 20sec). This would also have the advantage that I could tell the user not only why I flag his video. I could also tell him which exact part of the video I think is problematic.

>>4020 >The only thing I disagree with about the warns you called out: >I think just dismissing my entire thread in the way some of the posts do is equally non-constructive. I would argue that while I'm being ad-hominem, he's arguing "Argumentum ad lapidem" by dismissing the entire idea out of hand and not bringing any evidence to the table to support his ideas. Or, you are dismissing the criticisms out of hand because you don't know enough about ML to appreciate them. Do you want to try again from the top? How does your proposed algorithm differentiate an actual murder from an action movie scene from a comedy sketch involving blood? >But I haven't so much as heard a single tech term from you, or a real response to the thread that a liberal arts major couldn't have written. I don't need to use technical terms to explain to my little brother that you can't just write a computer program to win the lottery because computers don't work that way.

>>4038 > If lets say 20 seconds of the video are flag-worthy, then the whole video is flag-worthy. That gave me an interesting idea. What if there was a lighter analysis system on top of the final system to look for more broadly flaggable data, that pointed the final AI to the sections that need deep analysis? I could imagine a *much* more course grain system that lies on top of the final system, analyzing titles, comments, visitor traffic, most viewed sections of the video etc... to try and "set off" the main analysis system and limit the number/length of videos that are analyzed the hard way. >This would also have the advantage that I could tell the user not only why I flag his video. I could also tell him which exact part of the video I think is problematic. I was thinking that would be a wonderful result. You could also do the reverse in the user UI and encourage "offended" advertisers to flag sections of videos rather than the whole thing and use the explicitly flagged videos as training for the section analysis bot.

>>4048 >That gave me an interesting idea. What if there was a lighter analysis system on top of the final system to look for more broadly flaggable data, that pointed the final AI to the sections that need deep analysis? The Problem i See with this is that i would Run the Network every time a Video gets uploaded. Running inference on a neural Network is not expensive. Training is. At the point of upload, there is no such Data such as comments, view Statistics etc. For obvious reasons. I would argue that at the time where a User can See the Video, the damage has already been done. Maybe i should further explain that i would only use such a classification System for things like violence, nudity etc. I would not use it for things like illegal opinions, insults, exposure of private Information and so on. I think that these things are too difficult for todays machine learning. Sometimes these things have to be settled before a court. Furthermore laws are different per country. And also for such things i think the Information lies in the Audio not in the Video. Besides that, I would not want such a System to exist because of the Power it would give to oppressive governments.

>>4071 >The Problem i See with this is that i would Run the Network every time a Video gets uploaded It seems like Google does the scanning sometime after upload. I could imagine they'd have to given their thoroughput. >I would argue that at the time where a User can See the Video, the damage has already been done. That's true. >And also for such things i think the Information lies in the Audio not in the Video. Besides that, I would not want such a System to exist because of the Power it would give to oppressive governments. Reminds me of this contest I saw on Kaagle today: https://www.kaggle.com/c/passenger-screening-algorithm-challenge Could it be more Orwellian? I agree with that assessment for youtube too, but sadly they are already doing some kind of analysis, so I think it's more about magnitude of harm at this point. The current system already chills political speech on the platform. I just think it'd be more fair to do that in a way that let people know what they did. > the point of upload, there is no such Data such as comments, view Statistics etc. For obvious reasons. At least there would be a title and tags.

>>3963 Sure there are. Consistently beating AlphaGo at Go is something no human can currently do, but better ML algorithms can.

just use agnostic queries also, nothing is impossible unless you try to make it possible both deductively and inductively and then arrive at a contradiction or inconsistency which can be demonstrated and redemonstrated infinitely many times due to its reproducing steps having been properly sistematized under a universally agreed grammar