Deep Shit: Paradigms, Paranoia and Politics of Machine Intelligence
This is a modal window.
The media could not be loaded, either because the server or network failed or because the format is not supported.
Formal Metadata
| Title |
| |
| Title of Series | ||
| Number of Parts | 234 | |
| Author | ||
| License | CC Attribution - ShareAlike 3.0 Germany: You are free to use, adapt and copy, distribute and transmit the work or content in adapted or unchanged form for any legal purpose as long as the work is attributed to the author in the manner specified by the author or licensor and the work or content is shared also in adapted form only under the conditions of this | |
| Identifiers | 10.5446/32966 (DOI) | |
| Publisher | ||
| Release Date | ||
| Language |
Content Metadata
| Subject Area | ||
| Genre | ||
| Abstract |
|
00:00
Virtual machineInfinite conjugacy class propertyComputer animationJSONXMLUML
00:21
NeuroinformatikGroup actionArtificial neural networkInterpreter (computing)RamificationType theoryBitRussell, BertrandMetropolitan area networkComplex (psychology)Wiener, NorbertProcess (computing)Pattern languageCyberneticsAdventure gameMachine learningSequelGame theoryVirtual machineTranslation (relic)Lecture/Conference
02:04
MassHypermediaTraffic reportingAlpha (investment)Artificial neural networkGame theorySoftwareComplex (psychology)Event horizonWordBeat (acoustics)Meeting/Interview
02:34
Computer chessFrustrationGame theoryNeuroinformatikEvent horizonNumberVirtual machineAlpha (investment)Lecture/ConferenceMeeting/Interview
03:13
Game theoryQuicksortForm (programming)Alpha (investment)Pattern languageSoftwareVirtual machineLecture/Conference
04:00
NeuroinformatikPattern languageMedical imagingCoefficient of determinationLecture/Conference
04:27
Medical imagingGoogolForm (programming)Alpha (investment)Multiplication signVirtual machineOpen sourceSet (mathematics)Meeting/Interview
04:54
Process (computing)TensorComputer hardwareSoftwareGraphics processing unitComputer graphics (computer science)Lecture/ConferenceMeeting/Interview
05:28
Infinite conjugacy class propertyExecution unitCoprocessorProcess (computing)Pattern recognitionTelecommunicationMedical imagingNatural languageTask (computing)SoftwareGmailComputer hardwareWave packetLecture/ConferenceMeeting/Interview
06:03
InternetworkingWireless LANMetreProjective planeSuite (music)FacebookQuicksortMedical imagingSurfaceLecture/Conference
07:04
Grand Unified TheoryNumberFacebookInternetworkingVideoconferencingComputing platformLecture/Conference
07:40
Virtual machineExecution unitFactory (trading post)WordLecture/Conference
08:04
Directory serviceFacebookComputer simulationSoftwareBuildingStability theoryOrder of magnitudeMathematical analysisTerm (mathematics)Coefficient of determinationMedical imagingVirtual machineLecture/ConferenceMeeting/InterviewComputer animation
08:28
FacebookTask (computing)Coefficient of determinationVideo gameOnline helpComputer animationMeeting/Interview
08:50
Medical imagingDigital photographyCovering spaceSoftware testingNetwork topologyMultiplication signSlide ruleMeeting/Interview
09:13
Disk read-and-write headWordMeeting/Interview
09:36
Web pageDisk read-and-write headMeeting/Interview
09:58
Directory serviceFacebookExecution unitDivision (mathematics)Medical imagingDescriptive statisticsPhysical systemArtificial neural networkExtension (kinesiology)Asynchronous Transfer ModeWordMeeting/InterviewLecture/Conference
10:21
Semiconductor memoryConcentricVideoconferencingMeeting/InterviewLecture/Conference
10:48
Computer networkKolmogorov complexityFormal languageSymbol tableFacebookMathematical optimizationArmExtension (kinesiology)WordVector potentialMedical imagingAuditory maskingArtificial neural networkConnected spaceNeuroinformatikConvolutionLace2 (number)Meeting/InterviewLecture/Conference
12:20
Level (video gaming)CybersexNuclear spaceSummierbarkeitLecture/Conference
12:48
CybersexHypermediaPhysical systemProjective planeResultantLecture/Conference
13:34
InternetworkingSource codeInternet forumOperator (mathematics)Form (programming)Scaling (geometry)Lecture/ConferenceMeeting/Interview
14:03
InternetworkingEntire functionInternet der DingeMathematicsPhysical systemInformationLecture/Conference
14:31
Level (video gaming)Greatest elementHypermediaBit2 (number)PlastikkarteUsabilityOvalAuthenticationLecture/ConferenceComputer animation
15:47
PlastikkarteVotingDesign by contractSystem callForm (programming)Moment (mathematics)Lecture/ConferenceMeeting/Interview
16:17
Spectrum (functional analysis)OracleLecture/ConferenceMeeting/Interview
16:46
Self-organizationSeries (mathematics)BitComputer animationLecture/Conference
17:13
Inheritance (object-oriented programming)GodAlgorithmQuicksortField (computer science)Instance (computer science)Form (programming)Self-organizationLecture/Conference
17:45
Inheritance (object-oriented programming)LogicNeuroinformatikNumberBinary codeSoftware developerLecture/ConferenceMeeting/Interview
18:26
NeuroinformatikRow (database)Multiplication signOperator (mathematics)LeakSlide ruleRule of inferenceLecture/Conference
19:23
AreaTraffic reportingVideoconferencingConnected spaceSystem callLecture/ConferenceMeeting/Interview
20:03
Form (programming)Different (Kate Ryan album)Connected spaceTheoryBitObservational studyExecution unitTowerLecture/Conference
20:44
Connected spaceWordSign (mathematics)Operator (mathematics)Meeting/InterviewLecture/Conference
21:12
Group actionLecture/ConferenceMeeting/Interview
21:34
Virtual machineMachine learningTheoryLecture/ConferenceMeeting/Interview
21:56
Lecture/Conference
22:18
BitMultiplication signLecture/ConferenceMeeting/Interview
22:43
HypermediaField (computer science)Programmer (hardware)NeuroinformatikContext awarenessSoftware developerTheoryComputer scienceLecture/Conference
23:36
Row (database)MathematicsLecture/Conference
23:59
NumberData structurePoint (geometry)QuicksortVideo gameTouch typingLecture/Conference
24:41
Context awarenessPrice indexTerm (mathematics)Software developerMultiplication signElectric generatorUniverse (mathematics)ExistenceLatent heatDependent and independent variablesQuicksortPhysical systemScaling (geometry)Sound effectTowerEndliche ModelltheorieMeeting/InterviewLecture/Conference
26:35
CASE <Informatik>Multiplication signLecture/ConferenceComputer animation
Transcript: English(auto-generated)
00:17
talk. It's good to see that quite a few actually weren't
00:21
completely paranoid and disenchanted after Trevor's talk and left. I want to talk about something that is very closely connected to what Trevor was talking about when he spoke about images, but go a little bit deeper into the more general ramifications and infrastructures that underlie
00:42
the phenomenon of artificial intelligence, machine learning, also distributed computing, blockchain, and some other things. To do so, I will start with a letter from China that reached Norbert Wiener, the inventor of cybernetics in 1948,
01:03
I think. It came to him from one of his colleagues and friends called Yuan Ren Chao, a very interesting type who worked as the interpreter of Bertrand Russell and Albert Einstein during their travels in the 1920s in China. Norbert
01:23
Wiener met him when he was also being a guest professor at Tsinghua University in China in the 1930s. In this letter, Yuan Ren Chao, who was a linguist, who also, funny side story, did the first translation of Alice in
01:41
Wonderland by Lewis Carroll into Chinese, which became such a big hit that it spawned a Chinese sequel, Alice's Adventures in China, because what is even weirder than Wonderland? In this letter, he spoke about the game of Go as being, as you can see, something that might be of interest for computational processes and questions of
02:04
complexity. You might remember that last year, the game of Go appeared in the media when Google's software Alpha Go beat the Korean great-grandmaster of Go called Lee Se-dol. This is how he looked afterwards. In Korea,
02:24
this really resulted in a cultural shock. The Korean government the next day put a billion dollars into artificial intelligence research. There were newspaper reports about mass events of people getting completely hammered out of frustration. So you could see that this
02:43
somehow had an impact. And you remember, chess is something that has been done by machines for quite a while now, but with chess, people said, okay, it's a brute number game, basically. Once you have a computer that's strong enough, you don't really have a problem. Whereas Go was something that people always connected
03:02
with human intelligence, with intuition, with something that was genuinely human, and people were really puzzled by watching the game between Lee Se-dol and Alpha Go because Alpha Go did something that was unforeseen, that was unexpected, and therefore it led people to believe that this software had developed
03:22
something like intuition or even intelligence. This, of course, is not really true, but I will come to that later. Alpha Go was also something that wasn't really new. It's interesting to note that deep learning came to us in the form of bread and games, if you
03:42
will. You might remember these sorts of pictures that came out a couple of months earlier, Google Deep Dream, which also worked with enhancement of patterns that it finds. It's sort of, we thought that it's something like machine hallucination, which holds
04:01
true if you see hallucination as something that finds patterns somewhere and just enhances them and enhances them and enhances them. As Trevor has pointed out, these images are not images in the way that we are known to deal with images. These images are basically, the computer doesn't really care
04:22
if it's a dog or whatever. The image is a very interesting data set to work with. But for us, in form of images and games, these are images that don't show something but that obscure an infrastructure behind it, which is called TensorFlow, which is
04:41
the Google tool set for machine learning that, at the same time Alpha Go and Google Deep Dream happened, was released as open source. This sounds like a very nice move by Google, who always try to pretend that they're the good guys, but what it
05:01
did in reality was significantly enhance the user base and the experience that Google could gain from other people working with their tools. This went so far that Google was one of the first companies that completely restructured their infrastructure from software to hardware. This is the TPU, Trevor spoke
05:24
of GPUs, graphical processing units. This is the so-called Tensor processing unit, the first single-purpose deep learning processor that was built. So they completely restructured the whole infrastructure to feed into the deep learning
05:41
process. So every Google search you do, every image you upload, every communications you fire through Google mail is being used to train software and hardware to get better at certain tasks, like natural language processing, image recognition, or
06:02
whatever. Other companies followed suit. Intel acquired a company aptly called Nirvana, and also last year, all these companies came together to form a cartel called Partnership on AI. What could possibly go wrong? Another infrastructural
06:24
move that happened in the last couple of years was this campaign called Avila. Some of you might remember it. It was a big Facebook idea with the project internet.org where they were planning to build huge unmanned drones with a wingspan of 42
06:41
metres that would soar in the skies above the surface of developing countries, beaming wireless internet access down to the people living there, which, of course, also, again, seems like a very nice gesture, but what it does, I think there's very iconic imagery in this
07:01
image, this sort of really reminded me of this. So you can see the number here is a billion people getting access to the internet, which also means a billion people feeding more data into Facebook or whatever other platform. I'm going
07:24
to show you a little video about Facebook's self-perception of what their AI research is good for.
07:41
Our goal is to build intelligent machines to help people in their daily lives. What makes us intelligent is our ability to learn, and this is what we're trying to do with these machines. It's funny to note that the voice really sounds like one of those early machine voices
08:00
that you can use to read out word documents. To some extent, we get inspiration from biology. What we can do with current technology is build networks of simulated neurons on the order of magnitude of the brain of a mouse, let's say. So the way we train our machine, we show a bunch of images of, say, breeds of
08:23
dogs, and for each image we tell it what it is. We want the machine to be able to learn to perform the tasks. The more we show them examples, the better they get. And what's very interesting is that after a while it can do this with a new dog that's never seen before. It's able to exhibit superhuman
08:40
performance on a particular task. So how can we help people's life with this? Friends 937, news feed. Lindsay Russell updated her cover photo yesterday at 10 29 p.m. This image may contain outdoor, cloud, foliage, plant, tree.
09:01
This image may contain six people, child, close up, like one or more people, joy, smiling, 19 likes, 3 comments like Now I can see the picture in my head. Like, yeah, you shouldn't have been that close up. Like, now I can say it.
09:20
I love it. You have no idea. This is amazing. The whole saying of pictures being worth a thousand words, I think it's true, but unless you have somebody to describe it to you, even having like three words just helps flesh out all the details that I can't see. I feel like I can fit in.
09:41
There's more I can do. I can just call my mom like, yay, I've seen your picture and she'll be like, what? She's like, how you see it? Because my phone read it to me. It's new. And I'm going to mess with my mother head so much. In the future we'll be able to give a complete description of the image.
10:00
The two people happy? Are they sad? Are they holding hands with one another? It's this kind of comment that a human would make. We have systems in the lab that are able to do this and they will be deployed over the next few years. Artificial intelligence systems are going to be an extension of our brains the same way cars are an extension of our legs. They're not going to replace us.
10:21
They're going to amplify everything we do augmenting your memory, giving you instant knowledge and they're going to allow us to concentrate on doing things that are properly human.
10:44
There are a couple of things that I find interesting about this video. First of all, the incredibly biblical symbolism and language of making the blind see which, of course, again is a nice thing to do and if you think it a little bit further
11:01
and if you combine it with what Facebook is now working on or Elon Musk is working on with neural lace the creation of images on the neural cortex is something that is in the very graspable near future. Second of all they're all women who have to be helped
11:21
and third of all most of them are black so that's something to think about I guess. The fourth thing that I found interesting is the utter neglect for any kind of critical potential in there. He literally uses the word extensions
11:41
of our bodies in the very McLuhanian sense without even thinking about the political and critical implications that come with it. Yann LeCun has also been mentioned before in Trevor's talk. He has been around for ages. This is a paper he wrote in 1990
12:02
on optimal brain damage which is basically the approach now being used for computing in convolutional neural networks by pruning away redundant connections Just on a side note Yann LeCun of course always
12:20
reminds me of Jacques Lacan and the mirror stage so mirror neurons and mirror stage. Another kind of infrastructure I want to talk about is an actual older infrastructure or all the critical infrastructures that surround us one of them being for example
12:40
the Iranian nuclear infrastructure that a couple of years ago was as you probably know attacked by a very sophisticated cyber weapon the most sophisticated by far that was ever discovered and it took years it was called Stuxnet and it took years to find out that it was actually there
13:01
and then it took another couple of years to find out who did it so it was probably the Americans and the Israelis and this was a cyber weapon this was the result of a billion dollar project for many years that was a huge project that was focused on a very tiny thing like a little switch in a centrifuge that could sabotage
13:22
this system cyber warfare since then as you have heard recently it was a lot in the media has completely changed towards something that is probably still out there in the form of very sophisticated large scale operations by governments but this was a
13:43
forum post on a botnet forum last year where a user named Anna Senpai published the source code for a botnet called Mirai which turned out to be the thus far biggest botnet on the internet
14:00
which used a lot of unsecured IOT information on the internet of things devices to do such things as kick the entire country of Liberia off the internet shut down the heating system in Finland or do other things so I think that's a significant change in the way that infrastructures are being used
14:22
and attacked or infrastructures are starting to attack each other just to give you a short insight this is a threat map by a research company these are the botnet attacks that are happening just right now so you have
14:40
down here you can see infections per second I'm just going to leave it there in the background a little bit so you can see this is probably half of the devices in this room are currently involved in acts of war sabotage or simply
15:01
spamming something you get the idea the last thing I want to talk about
15:22
is another thing that has been very prominently featured on the media in the last one or two years after the rise of Bitcoin people started focusing more and more on the underlying technology called blockchain which is the idea that you can attach basically a crypto ledger
15:40
to any given entity which will then make it authenticable and usable for anything from smart contracts to voting and other things Vitalik Buterin, a 19 year old guy from the Ukraine formed an entity, you shouldn't call it a company, called Ethereum
16:02
which at the moment is booming quite a lot what I find interesting about this is that it was hailed by the left wing and the right wing in the same way this is something where technology again proved to be absolutely non-neutral
16:21
it somehow swung so far out into the left spectrum that it circled back into the right spectrum you had anarcho-capitalists praising it, you had alt-right people praising it you had any kinds of people praising it and then there is also the chance that it is being used by
16:41
the traditional financial industry or political establishment after Ethereum came this fierce looking entity called the DAO you can see these trigrams running around there you remember them from the South Korean the Bagua from the Yijing it also looks a little bit like
17:02
the Dharma initiative from the TV series lost, some of you might remember the DAO was the distributed autonomous organization that aimed to surpass any kind of other organization infrastructure by basing anything they could on the blockchain
17:21
thus completely taking out any form of human agency which also sounds nice at first but when you think about it gives instills a sort of notion of algorithmic agency that can be traced back to something like the idea of Adam Smith's invisible hand
17:42
or even back to the idea of the hand of God so something that is sort of super ideologic or super religious or super human in a way this is where the DAO comes from the DAO also meaning the way the trigrams themselves
18:02
help bring computing into being actually I'm going back very far into this early 18th century when Leibniz based on knowledge from China on the trigrams of the Yijing that were sent to him by a Jesuit missionary started developing
18:20
binary numbers thus laying the foundation for computing and computability as is and the DAO visits us again in these slides from the NSA from the Snowden leaks in tailored access operations
18:41
thank you very much there's time for questions that was my keyword by the way backstage, do we have questions from the audience?
19:02
so if you raise your hand I can see you, we have two microphones yes one in the first row maybe you can introduce yourself and then ask the question Hello I'm Meilter, I'm
19:20
working in deep learning since a little over a year now and I really have to say that there were a lot of falsehoods in your talk I'm not even going to name all of them but you just I don't think you have a very good grasp about what deep learning is about but then you also made some really weird connections
19:42
like what is the connection between DAO and TAO those are the two completely separate things and then also you said that in the Google video that the person was talking with a very artificial sounding voice but it's really just the voice of Yann LeCun who speaks French, so he just has a French accent I know, that was a joke
20:04
I honestly have to say it felt a little bit like spreading conspiracy theories here I sound like a conspiracy theorist? yes I didn't want to evoke the notion that I sound like a conspiracy theory first of all I was talking about not immediate connections
20:21
between different things like the DAO and TAO I was talking about, I come from cultural studies so we're apt to make connections between different things that are sometimes even metaphoric and I was looking at different forms of phenomena and technological infrastructures that are happening in the past years I tried to make clear as possible
20:42
how to how can you just say again what's the connection between the DAO and the TAO it's a so called signifier if you want it's the same word? it's the same word tailored access operations it's the same thing as the distributed
21:01
NGO there's just no connection there I was talking about different things and wasn't necessarily trying to make a strict connection between them there's no connection, it has nothing to do with each other those are just abbreviations that sound the same yes I know, I was talking about different things
21:21
I was giving different examples I tried to give you could see them as different chapters in a book I'm sorry if you disliked it I see a question in the middle of the audience over there, thank you
21:45
I'm also just following up on the machine learning and deep learning and KI and so on how much have you yourself tried to get into the theories of that because my perception is that there's
22:02
a great misunderstanding on what is actually possible and the speaker before has mentioned that and I've also seen that in other talks, so how much have you tried to get a grasp on what actually the technology can deliver
22:23
you mean how much have I actually worked with these things the sound is a little bit bad I'm having a hard time understanding you sorry, I didn't have the mic so how much have you tried to understand what these things can actually do and what they can't
22:42
as far as my capabilities allow me to I have a Masters in Computer Science which I acquired almost 10 years ago and I've never worked in the field so maybe we can talk later but I do my best to talk with researchers
23:02
and to talk with people like Trevor Paglen we did a publication together on the work that he is currently doing with a team of programmers on the visual aspects of it and otherwise I'm not a deep learning researcher, I'm not in the field of that, I'm a media theorist
23:21
and I'm trying to shine a critical light on developments that are currently happening usually trying basically to bring them to light and to raise awareness that there is something that we have to have a critical look on thank you do we have more questions? yes, one question
23:45
in the third row Hi there, I'm Eric, I'm a blockchain enthusiast and the thing I learned
24:01
today is as well being worried about the infrastructures which are creating our modern life and I sort of learned as well that I'm a blockchain enthusiast and I'm wondering or I'm learning through your talk that infrastructures are becoming more and more sort of not
24:21
necessary but as well we are losing the touch of where it's going to and where let's say evangelists are already heading up to and I'm sort of feeling that we are, I think we are like early adopters but our government is like a very very late adopter of those of understanding how technology works
24:40
I would like to know how you as well try to you said raise awareness okay, but how do you as well communicate with policy makers and governments in understanding what the possible indications could be
25:00
this is a huge problem I think, I mean, in general and I'm generalizing broadly of course, is that policy makers sometimes come from an older generation that tends to be affirmative in a certain kind of way and people who are early adopters of technologies come from
25:21
a very young generation who are very affirmative of certain developments without necessarily having the experience of thinking about long term effects of it there's not much in between so what has to be done and what still isn't being done on a massive scale is public education and this starts in
25:41
kindergarten and goes through the high school system and to universities and some sort of understanding that people have a responsibility as citizens and that too, because we still think that all these kinds of things that I talked about, even though
26:00
there were specific examples like Ethereum or the DAO or whatever who knows how long they're going to be around that these are things that do not happen outside their realm of existence but touch them so I mean, what I do is I teach
26:20
I teach at universities and I give lectures and I try to talk to people also in politics but I'm not a politician I think we have time for one last question one last question okay, that's not the case
26:42
then thank you Paul Feiglfeldt
Recommendations
Series of 3 media