Good Artificial Intelligence: Explainable and without Bias

Video in TIB AV-Portal: Good Artificial Intelligence: Explainable and without Bias

Formal Metadata

Title
Good Artificial Intelligence: Explainable and without Bias
Alternative Title
Good AI: On Bias, Interpretability and Robustness
Author
License
No Open Access License:
German copyright law applies. This film may be used for your own use but it may not be distributed via the internet or passed on to external parties.
Identifiers
Publisher
Release Date
2019
Language
English
Production Year
2019
Production Place
Dubrovnik, Croatia
Focus (optics) Presentation of a group Message passing Goodness of fit Machine learning Forschungszentrum Rossendorf Virtual machine Set (mathematics) Bit Quicksort
Slide rule Service (economics) Scaling (geometry) Multiplication sign Weight Interactive television Virtual machine Computer Bit rate Sphere Mereology Machine vision Power (physics) Mathematics Message passing Machine learning Personal digital assistant Googol Natural number Website Quicksort Arithmetic progression Task (computing)
Machine learning Dot product Regulator gene Decision theory Algorithm Virtual machine Computer Computer programming Computer Formal language Data model Machine learning Prediction Term (mathematics) Pattern language Data structure Task (computing)
Algorithm Distribution (mathematics) Decision theory Virtual machine Set (mathematics) Similarity (geometry) Function (mathematics) Student's t-test Computer Data model Goodness of fit Software testing Endliche Modelltheorie Traffic reporting Task (computing) Physical system Area Machine learning Distribution (mathematics) Decision theory Weight Sampling (statistics) Physicalism Word Grand Unified Theory Prediction Personal digital assistant output Software testing Pattern language Whiteboard Data structure
Machine learning Complex (psychology) Kolmogorov complexity Real number Virtual machine Mereology Sampling (statistics) Data model Personal digital assistant Function (mathematics) Vector space output Representation (politics) Endliche Modelltheorie Endliche Modelltheorie Object (grammar) Quicksort output Task (computing) Physical system Physical system Task (computing)
Codierung <Programmierung> PRINCE2 Multiplication sign Artificial neural network Virtual machine Online help Computer Code Dimensional analysis Word Medical imaging Mathematics Term (mathematics) Finitary relation Bus (computing) Hausdorff dimension Spacetime Metropolitan area network Personal identification number Metropolitan area network Touchscreen Dimensional analysis Voltmeter Line (geometry) Semantics (computer science) Similarity (geometry) Subject indexing Number Arithmetic mean Search engine (computing) Website Endliche Modelltheorie Representation (politics) Spacetime
Building Codierung <Programmierung> PRINCE2 Source code 1 (number) Dimensional analysis Evolutionarily stable strategy Dimensional analysis Similarity (geometry) Word Arithmetic mean Voting Term (mathematics) Vector space Vector graphics Website Einbettung <Mathematik> Quicksort Metropolitan area network Spacetime
Context awareness Building Gender Multiplication sign 40 (number) Archaeological field survey Analogy Mereology Perspective (visual) Dimensional analysis Area Word Vector space Finitary relation Hausdorff dimension Endliche Modelltheorie Office suite Physical system Metropolitan area network Source code Email Theory of relativity Hoax Sampling (statistics) Menu (computing) Formal language Website Endliche Modelltheorie Right angle Pattern language Representation (politics) Spacetime Computer file Artificial neural network Virtual machine Maxima and minima Event horizon Voting Profil (magazine) Term (mathematics) Googol Spacetime Game theory Metropolitan area network Task (computing) Standard deviation Semantics (computer science) Similarity (geometry) Mathematics Number Extreme programming Personal digital assistant Einbettung <Mathematik> Library (computing)
Metre Gender Multiplication sign Translation (relic) Control flow System call Revision control Mathematics Arithmetic mean Internetworking Extreme programming Search engine (computing) Personal digital assistant Internetworking Term (mathematics) Googol Website Right angle Musical ensemble Fingerprint Writing Software development kit Directed graph Spacetime
Email Internetworking Observational study Personal digital assistant Compass (drafting) Computer scientist Object (grammar) Physical system
NP-hard Multiplication sign Point (geometry) Mathematical analysis Compilation album Associative property Virtual machine Physical system
Point (geometry) Pattern recognition Presentation of a group Multiplication sign 40 (number) Graph coloring 2 (number) Wave packet Subset Term (mathematics) Endliche Modelltheorie Website Default (computer science) Default (computer science) Pattern recognition Matching (graph theory) Cellular automaton Forcing (mathematics) Point (geometry) Keyboard shortcut Parameter (computer programming) Cloud computing Database Thresholding (image processing) System call Similarity (geometry) Software Personal digital assistant Compilation album Website Right angle Game theory
Software engineering Digital electronics Real number Computer-generated imagery Source code Insertion loss Number Medical imaging Latent heat Causality Software testing No free lunch in search and optimization Fisher's exact test Position operator Physical system Pairwise comparison Distribution (mathematics) Prisoner's dilemma Data mining Error message File archiver MiniDisc Website Video game Endliche Modelltheorie Social class Software testing Right angle
Common Language Infrastructure Multiplication sign Coma Berenices Number 10 (number) Shooting method Film editing Error message Website Video game Endliche Modelltheorie Software testing Social class Fisher's exact test
Laptop Data model Film editing Integrated development environment Personal digital assistant Interior (topology) Maxima and minima Website Right angle Special unitary group Directed graph
Finite element method Different (Kate Ryan album) Information security Information security Fingerprint Local ring
Machine vision Pattern recognition Computer-generated imagery Software testing Open set
Mathematics Dot product Shooting method State of matter Video game Control flow Physical system Physical system
Machine learning Pattern recognition Confidence interval Digitizing Multiplication sign Computer-generated imagery Artificial neural network Sheaf (mathematics) Bit Mereology Benchmark Demoscene Number Medical imaging Coefficient of determination Fuzzy logic Software testing Endliche Modelltheorie Musical ensemble Physical system
Medical imaging Pattern recognition Personal digital assistant State of matter Personal digital assistant Website Computer network Range (statistics) Bit Pattern language Software testing Physical system
Ocean current Mathematics Module (mathematics) Website Sound effect Bit Physical system
Dot product Artificial neural network Computer network Perturbation theory Parameter (computer programming) Digital signal Perturbation theory Line (geometry) Mereology Field (computer science) Latent heat Software Website Social class Spacetime Right angle Endliche Modelltheorie output Extension (kinesiology) Physical system
Complex (psychology) Algorithm Virtual machine Artificial intelligence Control flow Insertion loss Complete metric space Instance (computer science) Data model Goodness of fit Endliche Modelltheorie Address space Position operator Physical system Regulator gene Decision theory Artificial neural network Computer scientist Instance (computer science) Personal digital assistant Function (mathematics) Sheaf (mathematics) Website Endliche Modelltheorie Right angle Object (grammar)
Axiom of choice Context awareness Functional (mathematics) Group action Decision theory Set (mathematics) Insertion loss Online help Mereology Wave packet Formal language Medical imaging Crash (computing) Mathematics Machine learning Natural number Endliche Modelltheorie output Extension (kinesiology) Loop (music) Physical system God Decision theory Sampling (statistics) Planning Division (mathematics) Bit Instance (computer science) Control flow System call Loop (music) Software Function (mathematics) Order (biology) Mathematical singularity output Website Right angle Family Directed graph
Shift operator Service (economics) Digital electronics Multiplication sign Digitizing View (database) Decision theory Virtual machine Planning Total S.A. Instance (computer science) Software maintenance Product (business) Wave packet Goodness of fit Bit rate Forschungszentrum Rossendorf Website Energy level Pattern language Endliche Modelltheorie Task (computing) Physical system Directed graph
what should we are stuck. the singer most complete what you mean it. i. the. all. good time to. the presentation of a on today i am also fuel obviously have hurt. and probably have done some work on a machine owning your message if not the sores on the left. the problem i will start from the beginning and i will not. given all of your machine learning techniques icing that spin or there will be a lady. presentations about that. i really have a focus on some issues which. also important besides the main the. so isn't the main message. we each focuses on what can go wrong in a sense to his machine than i thought it should not to tell you to use machine learning. the really good set of tools but it shouldn't make using a little bit before you use a i met sorts because some of these things which can go wrong actually can be avoided easily some more difficult to avoid of some. o.b. us once you know what you have to look at it. could it.
so. this is a site which you can go to in doing that which is kind of a bloke. on by somebody who has worked a lot in a and machine learning does so the business can be fully he was so the the act of microsoft asia for some time here are some m. responsible for a movie said. it should who ruled in china for some time and then he ordered his book which use it does think is a kind of into just to walk off for the eye. to come. i'm very optimistic years of a new way. on. bolt saying ok now i was seeing changes which will see what changes are but obviously it is told that. some of these are the sorts of performing better than we thought maybe twenty years ago and so should go there. it. find some posts which are some slides let's say which the which i find his way night so he says i'll the estimated all power to understand ourselves which is difficult i so we don't yet know how we actually seeing but to ten solid that. some of his the aisle gazans actually perform better than we sought. actually. performing better than some l.a. us like. and. um. like but unlike a nation that in humans want which is quite amazing action. this is so one side of the i'm calling and is also light to the.
um. to the vision or to the maybe if you said negatively the hype that this relationship was seeing real progress that will finally the be able to an get and diffusion intelligence which does not show. persepolis humans in some specific quite net oil heiress but actually which the actually came to what humans can do which is if we sing about it quite a lot. and this is still i'm a part of the ai height in so that the us today on and off into relating where we are no and moving at it in all the us which is. he was oh no ten starts to acknowledge which is not to write so you cannot scale up. i saw its in all of us in a sense that we get something which is as powerful us human mind. why is that the case and what can go wrong actually we. if we apply machine learning message if we tried to scale up the free trade a scale white in defend the us to discuss the debate let's look i'm a machine learning.
what we actually have found what we actually do without going into ago since machine learning in a sense is actually quite easy but was a powerful in the sense that we certainly and suddenly do not have to program computers to do some. specific task. other let alone but which is a powerful concept which is how we learn language show we live on the front. them. different tasks we just live on we don't we do not have to record i'm obviously and this is then tools of for computer who uses machine learning mess of snowfall dots these valk actually. ideas very simple. you have a lot of data. and in the sea.
lot of data and then you have to sell goes from which tries to analyze this data and which tries to recognize and find patterns and regulators and term these patterns these legal and it is sit term skype's in the met to kill him.
mathematical manner. which recall a model this model physical in up as and despair turns this like alone it is and this is it based on this model then we can in a way use these buttons to addicts things we get a new report then it's you. coming that depend on hold was afford a new import we can to predict that output for this specific important same way that the existing input from the data we have put into the cellular thousand and ten. this formed into outputs the students patents and then we can do the fencing so that no going again.
in some specific detail is a very simple example i hear we have ten ten examples ten samples of something where we have to ski at the butte send you see that it was should be. you should be enough to pick the wider on that out which should be tool for also it and so we use our learning years ago as some we train a model which basic kelly it was an sos these patterns from the soviet abuse to the. board might be some decisions he might be something more complicated and then we have his model and then wait for tested tight and we tested in the same way we give some examples some new example which we call the tests that you some new examples today earlier this month see how the modeling actually. it works with these new examples of what is important is that these two examples sets have to be the from the same distribution so somehow have to be similar i wish i let see peak. and on. i'm so upset here in this room. and they use is a painting said. the news that i still few words the test said that you are in some ways similar as the guards. the good of science or as the guts the students for this the foetus net to our q a friend in many cases and this is why it might fail because the distribution actually it's not elitist same year old coming from different to us. really this is a machine learning. some set up only works if you have painting in tests that coming from the same. area the same distribution. on. otherwise your loan patterns which on the tool for another set of examples however if you if you do it like that. you can actually get quite powerful systems and if we ask ourselves how would our legal system some models can become than we realize that for tasks will we sell says humans.
don't have to singularly we met some input to sum up would be like a nice an object the. understand a sentence i we don't have to sing about that if we have this kind of tasks then when we can build a machine learning system which is as good maybe even be better if and is as important useful model. complex enough to kept short clip wasn't all these similarities. if and it's always important if we have. enough import of what data is attending set and these him put out the data should be. centered even some way for the whole a set off examples which can imagine. it's the first seeing where something can go alongside maybe you have on. and that's what you're seen this picture this is the elephant and if you if you let's assume that when these issues such as they can only kind of us the part of the elephant like a nice the part of the elephant.
which they can touch but they don't see the elephant said stock and then his first one says ok this is a telephone book it seems like a peak snicket some. this is what would an elephant is this one actually says no known as the toll. this congress this is what an elephant disparate and so a body that are looks at an elephant in the us click good way to hide the last one says a case it showed a little mouse and i have it in my hands so it's obvious but all of them are wrong because all of them have. a convict and set off except for us but not that representative said and that is sort happened so obviously issue of data which not representative for focusing you want to learn or two percent in the model. nice think you have helped him in one of the talks about them but thinks and seeing not many of us actually looked at the technique and a real leader for gore little between buddhist technique.
i mean there's this idea of having care computer by a leading light so what we were to dual we read a lot of folks and then we know all. and that which denies in knew we'd have to. machine loan by eating for example eating or newspapers which which which we have. certainly the in a certain time based on this we get an understanding of the volt as it is and we get an understanding on housing's connect to each other. and how do we do with it. we basically learn how to sing still ate right and let's keep for what was the go ahead with a small example of what does it mean housing's comic nano i have here.
about ten concept site like man woman boy girl pins princess screen king morning which i want to understand. dues into an informational to resist them if i put it into such a change in what the search engine bus when it builds the index is just. i'm representing each term as it is and and if you search for men use your document what we have men but and this step is in taishan and it was sending a term by itself is called one halt and coding. because physically you use the thames as dimensions if we have fewer nine terms and play chelsea we say it's the term or not that so. in. one of the dimensions is one which is what does term is so it's its itself and us to see them. no that's ok if you want to do information appear if you want to search for terms in documents but it doesn't help you to understand ended or so is in some way. them quiet it done and you have a lot of seals dollars. senior won in in in each the image line now you can kind of too clever sings to. get away was this and not use much space site over the next ago as firms but you are not understanding in a single it because of term is the term. no one way to. twenty s. then how are these terms the ladies i'm is too. introduce some dimensions which kind of tide of character ice what the terms mean and here in a small example use the dimensions a username on the dimension male or female whether it's.
this is a male or female thousand which is expressed by the term use what i was younger old one of his business journal old and a certain mention whether. and this is a dispersed billups toil to you're not quite see dimensions and if i use this fee dimensional suddenly i can express the terms. by a footing. one source seals in to do so the dimensions. connecting the meaning of the dimensions two terms night so for example a man. is not she million man use not young try to your child and a man but so does not necessarily oil sort of all seals. i'm a princess. his female. the and belongs to a while to see ones and so i represent one of these terms by these the dimensions by saying ok this is how defeat into these dimensions this is a very clever idea. and this is in no way the ideal for but i'm betting same but here it is nine terms into this speed dimensional space i'm and this is the same. the this the this them makes the sings possible he tells me how come so relate to him and that's a powerful and sing.
i can the nile find that simulate term signed the same pot off these three dimensional space site. my can also use them. these space to the answer questions like an elegy question as a boy or girl is like it pains to whom i would find a princess. him and this is what we've been doing here but so does some this vote in buildings and does space where bet my terms of these sort of buildings allow me to.
sade use a cement he could be similar terms and this is how to relate to each other and mrs survey profile and india yield build embedding dollars more cross i do not only use the dimensions i use kind of like standard to five hundred and. and constructing speed of five hundred dimensions manually gets really difficult light so i try to do de soto medically. this is sort of i'm betting stool and they can strike is so you to five hundred dimensional space and put diverts in the right part of this space by analysing texts analyzing culverts i used to get her. the which are outside the vicinity off. but so you should have over which which has a certain vicinity if you cannot avert which is used in the same context third was the same vicinity read means the same right if you sing about kind automobile they are used in the same way so the vicinity looks those similar and the will be. same part of the space and. not the same but similar to a movie child a part but still. that's which go along was them to. a very similar and so this is the task which has been tight for quite some time in sochi forty s. to over fifty years when. researchers started to walk on the eye and is a difficult task and here we get it for feed is this actually perfect we have is. helgerson analyzing newspaper us lots of them newspaper articles and we get to cement dick space.
well. i ever. yes things which can go wrong and to them the one office sings which can i could nice if we have gained look at it is so dimension which. which physical it up as an sos male and female we realize ok fine we afford the stems here cars and son chap boy who which are kind of male old visitors files show same kind of a female however so because this is a dimension which. is used for old of arts and some hold events have now she's female of us as male dimension and in many cases this is not quite what you would put into and on toilet g.i. in. definition of a termite because like sample library and it's not necessarily a woman. and be a man or woman men's school takes this profession. imagination. can be man or woman. my wife we look at events in his bedroom buildings they also inverts which and on the site said we have to sixteen he. the boats we've shown in more than just mail space site we have philosopher skip market texts seeing some live if was a genius sites such as my side but it's not told that obviously. managing his and on our side we have stylists housekeeper set us us and while he does told that in our society. many of these professions are more female than male this is of course not deficient this is something which reflects our society and this is expected a problem because we gauge. a simple to do sell it goes from. and texts which you and older. you fill in some ways sold you saw the kind of perspective newspapers which will use this important new york times garden whatever they still reflect society and therefore this goes and his machine learning system it just takes the data and recognizes and expects. buttons and if you know our society will we find certain patterns then we find it in a model that and therefore the model is not the semantic space which he finds these contacts concepts but rather it's a semantic space which reflects how in. i was society or wind a society which is. the presented by this newspaper articles certain relations hold and that might be fine idea if we are way off that we have to be have enough that of course but then if we use it and we have to watch out but because for example if we.
music for translation right and translate secretary they were it's not clear what i was a few millimeters second to be translated into german german had this male and female home secretary than automatic elites kit kats the it gets translate to secure to win which is female. the second ten but because in my semantic space it was more to the female site. we choose not necessarily what i want to and not even be voted on look at the translation if i use such a feat like in a search engines are changes use feedback. what people want to see that the only issue to be just such this is what i do a sink if you have a year goal issue such was second to you get you get sick just like you get a certain kind of pictures that which do not reflect the deficient definitional say. it to a but the other. kauto or more in this case the stereotype that this it. but it is. but it. a that's true i mean the allied slight and so you should notice you you can of course conflict. but you have to know it and you have to conduct.
was a french call told me. in france they never use the term worldwide draft i mean use internet and the for a few translated why should insulated some time ago maybe those of you know. you translate into internet if you translate internet break into english you get internet so this war is long. and of course other sings so so that's actually year not to cure wrote.
and elegance some which tries to. in a very late whether a person should be an let let out of jail but based on the possibility was not his person will. we learn from them be involved in and out like i'm in the future or not but. and term we have few examples from me lunch e-mail person's and.
it's known that the that people are in the us is an example miss those left edge of his is so that the problem which we wanted to project his computer scientists in his case was that to me be an. white people are treated better than like people and soul. the solution i make an ngo isn't which is impartial which is based on all the data you get and you get an object dish and better the room. suggestion that which unfortunately it's not towards the famous compass case study this this company which built to do so in this system unfortunately we don't know it is programmed what we can look at the outcome.
if the system.
is this so quite famous a goes from this analysis is the subject in public or his association which talks about the bias induce a in the outcome of his out it was me who read on no hard evidence and facts we can find out what his pilots are not so for example.
this is sent to be said to people who left one is a white male who. in what was stated a lower risk for future time the woman was a tie it is for future crime and. but really.
when you should look at what they did and what did it after what turns out that this was just really that a suggestion that woman in this case was a time risk and did not reoffend the men didn't know you could say this is maybe tool for this case because obviously production or was. just probabilistic but it turns out that if you look at many of these examples and this is what the people who look to you you find out that to that again this impartial low as one which looked a lot of cases to train the model was biased. just and t. to men did white people but the than. then like people notice another fine example you can use a.
face recognition software by amazon which is court recognition as a cloud service site is is easy to use you just upload some examples. and then the train a model and then you get to predict if i'm model which for example you could mean you could turn the game in this case just use to find out where the two pictures to the present a simple us. the unsold watch the parade it or so in the match or politicians in the us in the congress and the singers in the senate which is twenty five cells and criminals and let's see if some criminal is sitting. in the congress warning. some came in a sitting in the congress but here we got along his ites because we got a lot of may choose actually we got twenty eight matches. twenty eight off a congressman right way then take a pool people in this is a criminal database which was on so do well not in the same people to look to be similar. in and interestingly it out or freeze twenty eight matches forty percent were of color twenty percent. i know so induce in in induce induce a force much as we had tried to person of color while in congress with only twenty percent of people have called it sold it means she was let go short cut. and then face a vicious often discuss the us less well. no muscle enough to have said ok you use to special it i don't use it to percent use ninety five percent under his eyes will be better. some. but that's beside the point eight percent was the default light so you sure don't know any seeing use just this creates oversight. you get lots of congress aren't so this is why you should understand what you're doing before using any of his own terms and second of cross this serve this proportion between call of people in the force macis and call of people in the us in the in the congress and that. that points to defect that certain subsets of people are. and not try to identify it. he says.
another. cink which is also related to what kind of data used to ten years and no so this is a comparison the offer of two systems sufficient a system and keep that was quite good systems that i could devise what is on the big picture was accepted it as a horse on his big. teacher. one which is fine and develop very well actually. if kind of the same position wallace and then a colleague of mine in in in billion look that while visualized what the nurses were looking at what kind of features that were using indus indies soon. in disc a sense or to the keep net goes from looked it look here i hear so this i can have for and is looking at the whole site says looks at the house which makes a lot of sense. now this official earliest and looks here i mean that's no loss on the left. apartheid but if you look closely you realize that this something here which says the the top kid he showed it so does the caption which says is from an archive about cause us and so somehow the area. it's million livestock a good looks into this column and if she sees the sculpture which of course it doesn't understand what the outcome should be a horse site which is cool it seems for detaining circuit yet off cross this is an example where delegates and learn something. which is just too in attaining said but not to intervene real world solution to come was because of was as they are classified only once they are not from this is a specific a cliff and do not have his caption it. but for the us and for the test seems that the test images were kind of again the same distribution rights or that this find the source of the horses came again from this loss of life but it did not to prison the whole world in a way.
it. but it. it. so. a lot. so do you spell village in l a number of us.
this shoot the by having too many of whose costs to have pictures us so i'm an issue. the only known horse. you should put towards pictures which are not from his own hype didn't do it could under seal hundred years. here. the. now this is the who because i left the site also now look more like us the intellect times looked like us the number be nice cars which have a really nice and to they don't do anything but which is on purpose so at this is way off.
why was constructed likely like looks like you're thirty b.n. and and they are driving around and been going on in california it. loving miles and miles and tens of thousands and millions of miles. and this is not because people and he said just wanted to test those was an induced coma self telling us him but because they want to discard to alone and look at why meant look at what happens to it because this is not a people come to light. alone in copyright and the more lives the bed to get so it and therefore be lifted times for a long time in california. the ten million miles seven million miles and his works have a well if you let the cut life in california if you'd she moved on a per site it will crush.
because even though we did ten thousand miles in a ten million miles in in california night if you go to naples the a petri dish is a bit different and and and so big data is not necessarily representative date so this is what is the pictures us and. it also depends what you see on the from the verbal tight so this is the you can buy it has lied to explore brilliant has this out to pilot you should leave your hands on the steering wheel that because sometimes it does look quite book why does it not quite to so it's not good advanced but the.
also i'm one of the first accidents was actually this one where the car when it. a kind of under a talk right so just like was possible to devote it should not have done that but it's another meant to happen sometime site and the car which was sold diving which had. which it. which will sensing the environment and by comma went under the talk of crosses question but because to know why did that happen. she looked at the talks in the us they hear right so they don't have any sink in uk fix the would you have this kind of. i'm sheet site so if you are paid a sense a. and you have phoned in the uk should look around you cannot really looks below the cut the talk that you recognize there's a talk of the few in the us. you look a fool because the census data low and then you have to fall back on the camera nobody should have said ok fuel now using only the. no then you have to use to come up but in this case the camera was kind of handicapped because that was very sunny outside the camera could not really see the light so if you look into the sun you don't see anything because of this committee not see anything and the talk was a light truck. so we all look the same so did not see the talk. and the come and the dozens of the notebook and so in this case a trench located is also the says. stuff costs.
but as they actually a nice them but soon the port which was. which was just recently done in spring in much worse by a security company in in in china ten cent security and its look at what they're doing that's even the six know.
an. i. ok that to stiffen to its.
it. i.
it.
each shoe shining lights which doesn't matter if we don't understand things let's look at it.
so the best thing that has local pilot in different.
situations.
i.
the test for example of the white first wife was welcome at the feet of pain side day.
and a start which is useful.
i.
ok this is an opening.
you project a picture on the come on to become all too so system. and it seems like a say in english stop and look at it in what is the so does not work right.
you have this strange picture. she's just not so but but one important singers have caused following lane side you should follow lanes you should have cost them.
i watch.
if so can you.
it is so lame following behavior is clear.
the slain shoot your state had quite a few dots on the on the street but let it on.
and somehow use more dots stop a system that and the goals of life.
it's all i do.
but come some sing way you can actually.
it's almost like it's ok he had a short holiday can steer using do so the most which is another scene.
let's look at what happens year.
and let's look at his first the in not have said a fixed number so we know already that on some benchmarks the legal systems image recognition know some so really really good better than humans so did make a nice digits you've moved a bit strange shy. but. so this is a sealed is a city looks quite sane should use a kind of fuzzy. science i still it was things like a bust perfect we know we change it to beat which changed to beat the left is still c.e.o. and eight get some dogs are sprinkled around. some and suddenly the system says ok notice is not a ceiling the most a five four instead of the season eight. he is even more dramatic saudis do science look very similar but i mean i'm human looking at it and moved so far nothing has changed from but the system says ok now you can go ahead this is the fee to go. or here you have to watch out sleepily and here you have to stop so totally on you should have the system in your car section not really good time. even more funny picture this is not funny this is dangerous and this is funny and it so they left the side is a panda it's not illegal to see because you just see part of the band as solar system is not an issue it says fifty seven percent confidence that should be opened up ok so actually blocking while it.
then it changed a little bit was this strange pattern which is invisible to see for humans that so their height site is still the panda but now the system says yes i know it is a given night and i'm totally sure that so that's really strange show that this is not just one system you can do it is. it was quite a few different systems.
and this is what happened was that this loud quite a lot of it. i hope now assessing all kind of our case so so what's called here so to have to squint of state of image recognition that back to just use the picture and test to find out whether it's raining or not.
and what people do is scanned of the this is their right de cock picture on the left you change a little bit in a way which humans actually do not know to sites or you tool and of the various more changes in their right side to still looks the same but she again this them. the situation it suddenly turns out that the left says the rains cause zero zero one sold as no vain. changing a little bit still looks the same it says sorkin pain scores point eight which is almost one so staining turn on the swipe us in the lane following behavior is a bit more difficult to take because the tesla's really good in recognizing lanes.
you can change that each of the pictures so that the kind of misses the lane so there's still a lane here but the. about the kind of vanishes. you should try to do it in reality it sold on the left you put a lot of thoughts on this left lane is even more visible light but it has the effect for the system that it says a kid s. the lane anymore year but this is actually very difficult to two. to use because in a way to for humans is obvious that days the lane i like of nies it immediately the that the that this system is doing something wrong current but turn.
but what you can do easily is kind of to simulate lanes will go no lane site so the tesla's very keen in finding lane site and this is what you can what you can what you can use you would use small dots on the left side which you don't see the wall which almost no. sea and the and the tesla sees the lane and and this is what was exploited tight and sold this is what is happening key a that own.
examples which generated when my while you have in mind that that the that you want to tax the system so the there's a field and a day day to day import in in many many parliament us to have a small extent. but all are dues kind of perturbations but because many about of missions to a smaller one so we don't see them but on his part of patients at up in the illinois and and the us and crosses the line crosses the line from one careless the tool to dead like loss and. they are not very specific to wear particular. on network built across different networks. and and and so in a sense of tense out that is deep models and not that deep and not so deep and we saw quite so they do have. some say locally leno border site which which are not tool rights or day learn something and on or awake some first of a violent but to but if we change or sampras induce way they they feel ok so let's conclude.
what we learnt what i write to him playing across to this position was that we should not just use of us and obviously you should understand them we should always assumed about some some aspects of these addresses which. john not connected to the simple performance as usual but the which are connected to explain nation's story that is why do you suggest that person should tell me he's insight wide he died not get a credit. then the person or the human who uses those who should be able to say ok you didn't get a candidate because. i'm the cross whom you do not a break your last loan or whatever hype so they should be an explanation not at all park system like. it often is the case if the user just the neural network was organising additional that there should be no discrimination swelling awesome sight and the. and we saw his six from for in the complex case site because it's another goes from it doesn't mean it's more an object to phone than a human and we do not want peace and actually we are not allowed to do does that a loss which which say this is not a lot. and so as the computer scientists we can us think about the problem and should sing about the problem how to explain the in the such analogous i'm so be difficult on the model layers to modernise don't come.
the kid might be difficult to explain but for specific instances it's what is your right so i can explain why you did not a good said they cannot explain the model works for everybody but i can explain for one example i want the. this machine learning lessons to be bias fee or to say my data are biased and show you get on his site so you get by us there's arts this is an important question will send the want them to be to be a bust because we know into lots of high cost of the and of course a. we don't want to have an older person which kind of feels too easily fight take it and then for example if we are loyal then the question like some have become so what kind of regulations should we would if any on in.
containing data that them all should as select training to to shoot the early to use if should be inhuman in the loop some will not should be allowed off the beaten insert the us who is responsible if something goes wrong idea. if his car crashes side whose the to be blamed tried to copy the manufacturer were mean i know it's the human who drives the car but because they say you should not rely totally on the nose and white so it's the tireless days the sponsor for your family self. house becomes the front and all of this can be may be summarised in does nice picture which some time soon.
skype so how people use machine learning owners from sites or enough data collection which is to help it up people sensing which is difficult and you do model training and then you have just munich and you get the desired know me a decrease in the eye and you should not believe him then shoot it to understand what happens and. that's the choice thank you very much. oh and the question's commence s. so. the. well i. but they are the way it. was it. but. and so on. you are or what. much of it. also for the first mission to you can have caused it. but you might not be perfect nor is this but you should get people home or what have all but which i guess they did off to watch then you get but it's interesting to note that or so we as humans are not quite good in the in finding the diversity in faces which have. from somewhere else like it because or so our own of obtaining this is stained on a certain kind of like some purse. an. and the may be the best answer here is to say don't rely totally on the system and and be aware that some things can go on it so you try to be as tough as intuitive as possibly in your in your training kicks some both sites already very very important. and then you probably get better even in human and i am but still if you get really strange. in example don't sing school yeah i got into said takes a. this is still a bit and clia hall and what is the best way to confident because the said texas something different the exploiting call your pain does networks like it so physically does not to accept faint by using his function which says the cane reached auction house. thus the in the us the suffering can change but change more quite so you goin you do can some kind of fuel climbing try to go in a specific action. and it's two hundred changed out what. if you find input and lead stone god what you look at the straining function and said ok after changing to beat and and or because you don't want to change too much according to the nature of changes most adult court. while changing not very much in a coat so you are. from the problem was the sick division functions is that the that you you at order small changes. in and people have tried to to train was kind of changed except for site so you can take his to this to take samples of put the most important say what you what you should you get and that talks to certain extent but it does not believe that the problem some hollis the model not. but not attending process and the. it's not yet to complete solution to this problem. but. so. but so. it will. it's too so that's still the interesting is such a low. and yet it. i think it. it. it. if it were. the u.. and it. you. you. but . but you also want to help the people so the system was detained on this and is that a sense that if you bought this data sets be aware that it might not work well all about what the book at all so that would already be help for so finding out that something can go on solar. and children own or so more diverse site so many should learn a new language sure you realize that if you want to understand children is what if you have done understanding the out so that its to the public want. yes. what is. what makes it set out. it was. these will go so now it's see and i'm so so so easy. home on me. besides you might want and sure that the but it's the only. i switched. doubles the. but which have it explained. i always just he was gay and it's we who will last one which this. so so seeing you saw him that i'm us your so quick to point out is easier to explain such instances and not the whole model of its oil is used to explain why discuss a picture which shows lung cancer or not to mock on not. wall why does so i'm really listen to know is this a horse and how. so you don't explain our model you explain why this particular example i do get his and his suggestion on classification and that the xo rex already quite well i'm still and open to some open issues but the trucks well. whether you get perfect expand ability of on seeing so because we cannot even explain why we make physicians so it sewn in them. or so the doctor can not always say exactly why he came to dislike knows as such he can explain it to some extent often watson is actually funny she look at psychologically touch or that the pain is actually very good in explaining sinks diet. pos talk but so it makes a decision and consciously and then it explains why it has done just decision which is a national is asian of something which was unconscious only your nationalisation is they should not how you got his decision to show us the costs talk explanation that induces. what we might be able to achieve sheer that is interesting experiment side where you cut and some patients have the and you have these two parts of your pain and are connected if you cut it. because because of some illness also than your left lane doesn't know what to hide plane knows that and then you look at a picture was one part of your pain and you tube explanation was that apart and and it's totally and connected kites or the lift pain explained. this some sink which we should on which it sinks the right hand basket was just just comes up with an image. so what now. the us have. if it. but.
but on. now that way. yet. so. but. but. so. so. which not. it's that it's. the a rate that night. and it drives a national election of it. and. so i actually know what is so so q i'm doing because this is what they could sink. the total is important point. one in ten years so the new site will be gone to it and then we would have products which rule book while but they will not be magic. current them in the current situation as seeing you mean sometimes you are forced to argue against this kind of view i do some younger side this system which can do it wasn't because of and people tend to use it on leaders said he says. the magic i just put in a few weeks so so one one makes one sink wrexham produce if you do some issues some a commission task. an and you sure i don't know mechanical engineer whatever the. dan that ideas circuit so for example. finding out whether machine place but so sold his coat. so a shift to speak machines which would show used in production. the service every year level how few two years to make sure that they don't think you don't know exactly when they play golf course so you sos them more often and s. which wastes money and time sold as is elf preventive maintenance which tries to. to find out where the machine is going to take in in your food shujaat i'd and then uses so this is the task ahead so find out whether machine takes before i'd play excite locally enough that you can so is it and otherwise you don't suppose it's actually important task of talks is perfect but it's. lee hout to implement because these missions usually do not. the good especially the german machines that poor mention site they pay a once in a few years site so where do you get the dining data that the machine is going to play you don't have enough training did it so people are kind of collect or a or or send sensor data from his machine for five. your side and almost all all of the time it does not play tight and then you have one or two example where his body was playing and then and then people seeing ok because f.t.c. i get loans when it takes is not tow you have no examples that. question you can ask then to see such a person is would you as human looking at the data let's let's assume you can do to them and we are slowing looking at the top but in principle i would you find a pattern. which would allow you to predict what are the machine is going to play but if you're seeing a bargain if just one example you can immediately see no seem possible hide i have one example i cannot generalize it too but so that's a good question to find out whether you have enough data about what you want to learn. so i'm the debts that some seeing you can do to kind of and move away the magic a little bit and say ok she was a human not able to come up so if you have a solution like why should those and boots. in plain awaits fast the budget simpler than you. yes but the. the. what it. it all so it all. . but no noise. depends on the it depends on the same. the it depends on the. the education light so if your doctor in medicine explain it really does very important to it because the because as a patient you want to know why diagnosis medicine dr using such a system you want to understand why the system gives this suggestion i because otherwise you live or. so explanation civilian really important. you should rule of mission of digits it's not so important you want explanation if something goes on but usually you don't kill you just uses them as so many of the kitchen and. are you. if it. so what. here. the. well. i just wrote. and that was right. that's what happened. so you like what's that. it was. on soul so so one way to get explained ability. and his kind of the two feet a simple a model to york some fight so you might have a complicated model which is necessary to hand lowered examples you saw it and afar. by definition it's very difficult to explain as a model of it and you were to kind of like your plane thus you you you come up with the six when nationwide which we choose as of it is simpler model which is just tore not foyer sing it but his tool. what is particularly excited for and said this is this very simple decision to be which says it does and does his tool then this is the outcome and of course that's a pos talks explanation i have a much more complicated model but it's sufficient to explain certain instances so that's one one one possibility. to to pool to do so at them so use a perfect complex model get high performance and then use a supermodel to explain such cases. it. but it. it would be it. are there more so. you are so it's what was it that today. it it go a few but workers about that. will it go. that means that all the cool but who knows the disease and. and it definitely so also i mean you should start being a way out that your model is restricted by your training said which you sing about his obvious site but that too many people don't sing about it and then you can have tried to extend to attaining. set your examples of your tested and orzo issue for example me be managed to to kind of implement some kind of batteries inning had which sits atop a model and kind of like a nice as if you go out of my own side you should use them now is in a situation. and do well which is not covered by the pilot training examples side that's all it is a good solution years differently. that. so what. but that. only one.
what about me.
a separate or not.
Loading...
Feedback

Timings

  506 ms - page object

Version

AV-Portal 3.21.3 (19e43a18c8aa08bcbdf3e35b975c18acb737c630)
hidden