 very happy to introduce the next talk which joins two topics I find quite interesting for one thing machine learning in particular deep learning and for the other thing sustainability and how this can be connected and if probably the deep learning hype is probably quite enlarged which Natja Gaisler and Benjamin Hedash will discuss for us so let me stop talking here let's enjoy this talk give a warm welcome to Natja Gaisler and Benjamin Hedash very necessary we can be here and we're very excited to talk about this on our very first congress actually I'm Natja and I just finished the master my master studies at the TU Darmstadt and will probably start my PhD studies there I'm doing this for two years I'm Benjamin and I don't just want to think about how exciting deep learning is and how to apply it but instead I want to talk about its implications this talk came about by the topic of the 3063 with sustainability and so we always find exciting to see how machine learning and deep learning interface with sustainability so that's exactly what we were already busy doing because sustainability is currently just very topical and exciting so before we talk about three the three different layers of sustainability let's talk about the societal and environmental environmental impact of course we first have to talk about what actually do we mean by deep learning in these next 45 minutes in this talk what do you need to know so we can get you along in this talk how does it work on an intuitive level not so much on a detailed technical level and how is it used currently so when we're talking about deep learning we primarily mean this construction of neural nets artificial neural networks these are machine learning constructs which existed for quite a while and then interest in them dropped quite a bit because they didn't have the applications one was hoping for and currently they're actually again hyped and quite and quite a lot of use so we're talking about lots of connections between different nodes and they can be connected with different mathematical functions and each of these only represents a nonlinear application function and once the way it's between these calculations have been trained completely then each numerical input into this net gives you a specified output and this might give you classifications weightings which means for neural nets the most non trivial implication is that we need lots of label data so usually if you're working with neural nets you need lots and lots of data they need to be very diverse should really represent your reality whatever that means and like that you're training the model so we're seeing quite well in this case reading from left to right you're having input data which might be in some format which we will not discuss today these are going into the neurons the nodes and each of these nodes represents some specific feature so in image recognition this might be an edge or a curve if the further that you get into the neural net it will represent more composite and more complex properties and the trouble is usually we don't know quite well what the neural net is actually learning so we find an image classifier which can distinguish dogs and wolves but we cannot look into the neurons and see okay this neuron classifies what the tail looks like that's not how it works usually so we have black box models and in detail we don't really know what we actually learned in the neural net so we don't know how the output actually comes from the input but this is the basis for our systems which we usually talk about when we mean deep learning so what we have in this case is we use math lots and lots of data and use a few tricks from statistics so we use that when we look at things often enough then there is some system in them and if you look at them often enough you will find the system so it will generalize so this uses tricks which we know for hundreds of years and statistics and these are applied to get from a heap of data which we don't really understand and just use this the fact that it's lots of data to generate some generalization and systematization of this data at least that's the optimistic upshot so this has though it's called deep learning the implication so it sounds like intelligence and artificial intelligence so it won't really represent what you think of when you think intelligence and this year summarizes it quite well it usually yeah people usually say the human brain works on habit and conditioning much faster so we don't usually need to be hit 200 times in the face to learn that we don't like that but if we show a neural net 200 dogs and 200 wolves the system will not have learned to distinguish them so far because that it won't learn from context as the human brain does so human brain does different things to get a decision so we will talk about that in more detail later but the problem is we just go for a mass of data and we cannot reach the can not be as precise as we want to and originally the plan was to rebuild the brain because we have neurons that fire and kind of was the idea but that's not how neural networks work to die we don't rebuild brains it was just the original idea so the interesting part for us is how do we apply this technology we don't see this technology only in research or in universities but the technology is very far distributed and we see it everywhere after this bigger draw we talked about earlier this technology has quite a high at the moment and there's really a lot possible at the moment so there's speech assistance Siri Amazon they need to know how speech recognition and text and they have to process information from this huge information cluster but there's also companies like Tesla and Uber and through them there's like a lot of distribution of autonomous driving and they have to work with image recognition and stuff like that and so if you think about that it's quite easy to see that there's lots of those recommendation systems in everyday life that generate recommendations like Google and Amazon and you have like those rankings and there's also the question what do I get to see what does my Facebook page show me who gets to see what and that's not as straightforward as many people think a lot less known it's like systems like systems that do scoring for legal reasons or face recognition identifying people there's also scoring algorithms for some social systems for example for insurances and also application can be happening by those neural networks without any person looking at the application and these systems again work in a way that we don't really understand what's happening we actually don't understand what's happening for example face recognition or those application systems there will be a five-second video analysation and the system will do quite a lot like measuring distance between eyes and nose and then it might find it kind of looks like genetics and the systems are applied today and and if you actually look at them precisely in more detail and look what they're actually doing it there's quite a lot going on and there's actually a lot of disadvantages and they are biased and it's getting more and so now we want to talk about scientific scientific accessibility so we would check how relevant is the topic that we do how relevant are the results that we have for the field of study that we have and for the everyday life of people so we also questioning can we reproduce these results at all can anyone reading this paper really reproduce the data or even get at the scale of data so do we have published enough detail and are the results as valid to make these things possible so and can we reuse things that we produce in size or are they only just a one-trick pony for that one very thing so also are they competitive are there some systems that that are better with less effort so and we would like to check with what systems thinking is behind the system how did we research what is useful or not useful and finally what's the validity what's the impact of the things that we have as a result there so is it relevant is it statistically significant what we produce there so well at this point we let's have a look so I mean some of you might be right work in science some from industry but it doesn't matter so how would we wish that science would work so very systematically so people think about things they test things and I see they work and then it's fine but in reality it's very very common that we have a completely different model behind there so there are publications about some about how the neural networks work and how to how to design them how to make the data flow people think about some things they publish it and other things other people think well that sounds interesting let's just take this and build something for my use case but the new system so they just take a model that they heard somewhere that just was part of us popular in in science and they think okay how could we use that how can we take it and yes well I just put a couple of layers of those behind each other let's say just some ballpark number and now let's let's use a couple of dimensions for our vector they did just make it up it sounds plausible then they look at that the data you beat the data as long as they until they fit into the model then you do some numerical computing on then and finally you put it into the network and then it's called deep learning because that means now the learning starts so you push push the data into the system see how it fits how good is the prediction and looking at that then you refit the system you edit the system and you try and try again until you have you do some waiting in the functions and then you have basically have made up model that fits your use case if the numbers look quite good on the data you put into that then people say yeah well okay let's put this into the paper and they say okay for the classification of wolf is against dogs we have this architecture and those daytime and these are the numbers here you are this is our great new the research result if the data don't look as good they say well I probably use just the wrong system that someone used maybe I have no neural layer too little or wrong dimensions well okay then I just start again with some arbitrary thing it's just a power and time and so I let my GPUs run and run and run until they heat up and I start just just that over and see if this time the numbers are right and well depending on how this works you either say okay I use it or I do it again and in addition if you look at it it's it's not really a scientific it's not even empiric it's just try and error hoping for something good to come out of this but after that you can use those beautifications fairy dust methods that we have and science and these are that are hard to track for people who are not insiders so for example you just show the results where the data is fit and then the second data is it where the numbers don't fit as well and that don't work as much well I just don't publish those I just keep them in my drawer and and my numbers that look good I just publish those and someone just has to replicate those data and make those numbers look as nice as they are so and in many fields of study this is something that happens if you publish a paper at the prestigious conference with a minimal extension to the state of the art that's already something people are proud of but what you also can do is you can just repeat experiments and have the mean of those experiments you also can use that very experiment where you had the best score and just publish that and there's other tricks in those so we already have bad process that is misused to get better looking results and a much shorter time and to publish those and that's what we see a lot in the deep learning field of study not in all papers the fundamental papers are pretty well but a lot of application papers are actually really I mean the way they are made they don't have any additional value or only relevance to actual research this thing is obviously written in a very marketing kind of way to emphasize this a lot so just imagine there's this field with such pressure and so many reasons for stuff to fail or on the other hand to beautify yourself then you will obviously use this obviously we see it's deep learning is particularly exposed to these tricks for one thing so so there is research to understand these black box models and understand why they work the way they do but if you think of these as black box models is obviously making it more exposed for such tricks so the data is processed in some way and it's going into the system and then they are post-processed and then you have to decide what's actually true is this true enough is this okay to publish in my paper so what they actually measured will probably not be in the paper so this is actually where you can fudge a few things and to get it into your paper so because there's a huge demand for experts in this field it's actually quite easy to get through with methods like this on the other hand if you want to get the good jobs you need to have publications like this so there's a demand for such publications and there's lots of low hanging fruit in this field so you don't need to have much of a night much of good ideas much of brilliant or new ideas so you can build something quite easily accessible and you just happen to be the first one so you don't need to prove yourself in comparison to anyone else you just show I'm the first one to do this with exact liberal results so you're the first person to do this and then you can publish so that's why many people want to publish lots of things quickly in these fields so when we look into into how good a system is which was presented then it would be nice if we could just reproduce the experiments that's actually sadly not as trivial as you might think so even if the systems which are used are usually quite standardized and usually open source that's typically not true for all the details and modifications the people who publish the paper have used this will probably not be true so you will probably not have access to the evaluation typically you won't have access to the data because this is usually quite precious to have so you don't want to give it publicly but that's on the other hand not the way science is supposed to work you can just you can't just give you a system but not give the data my system is just excellent I can't give you the data to prove it in these systems there's loads of hyper parameters which you firstly usually guess and then you just do trial and error and if you don't know these hyper parameters you have actually no no shot at rebuilding this system you don't know with what separation of the data this has happened so you don't have a chance how to build the system but we need all these exact exact values to build these systems because they are actually quite fragile with regards to these parameters so if you just change a bit of the things about these chain functions there you get bad results and you're not sure if it's because of the original bad publication of it's you building the wrong thing currently there's a few streams to make this the duty of the researchers to publish things along with this but yeah there's nothing forcing them to do that but it's not obligatory to do so so imagine you want to have a seal to say my paper is reproducible then you have to give all this which we have on the slides of code data hyper parameters the random initialization the sequence and the grouping of data but you can just not do it so this is obviously a position where you want to have a discussion because yeah it would be good if we had this on all papers so the effect of all this is we have lots of research which cannot be used by other people so others have to repeat this research and additionally through this effect and the pressure of publish or perish research will just be published with only minimal optimizations so this system is better than 0.005 percent than the reference system and that's what we know the paper so we might wish for having reproducibility everywhere so the most important thing whenever we're sitting in such whenever we're sitting in such publicity we need to show this but now look a squirrel so that's something that's happening quite often because research even if they want to publish their source code they might just not have the time they haven't cleaned up their code by the timeline because the pressure to publish is so high it's raising exponential you have to be faster to be state-of-the-art to bring your own improvements onto the market so there's working the working quality doesn't improve with that so then I have code that's quite messy and then there would be much work involved to clean this code up to publish it and so everything else is more interesting than making this code accessible for everyone so of course that's not happening everywhere there's people who actually try to do that and put work into that but it's quite rare and so there was research this reproducibility challenge there were researchers were asked to look for one paper from 2018 or 2019 and to reproduce all the outcomes and they were asked to the authors were asked to publish their code and then other researchers were asked to reproduce this work and the code and here the success rate so there is at least 30% of reproducible work and 54% of some what reproducible work and then there's also the part of the difficulty so there is like reasonable difficulty but if it's very difficult to rebuild this code it might not be worth the effort and so with at least 20% of the papers it was really hard to reproduce the outcome and it was very easy only with a really small percentage and the question was do we have a reproducibility crisis in machine learning and people were asked whether they have a problem with this if they think there's a problem and just by doing this research the asked researchers answered 15% more likely that there is a problem another example so there are papers actually with which are dealing with that question how reproducible are other papers so they're giving recommendations like they're looking for the top-end recommendations so there were 18 publications from on deep learning on big conferences and people looked at them and looked how many of them can reproduce we asked the authors whether they can give us the code ask again try to make it work try to work with similar hardware and rebuild those systems and so for this example from 18 papers they could reproduce seven papers so for those seven papers they can rebuild everything and come to a similar conclusion but important only after they worked on it after they asked for it after they put into work to ask and rebuild it and that's not the standard so normally if I just publish a paper on a conference then people just read that paper and maybe they watch a video and maybe there's some additional data but normally people only read this paper there was 8 10 maybe 12 pages and then the people only decide with this text only with the numbers in the text the authors gave and then people decide whether this work is relevant important usable and then there is a decision whether that paper is published or not but there's people cannot actually test whether that's right or not they have to trust the text that's the standard if we don't explicitly try to reproduce those papers or ask for reproducibility on this great big conferences and there's not one of those big conferences that really require reproducibility it's just like optional requirement and the reviewing process is without actually reproducing the work and to make it a bit more demotivational out of those seven outcomes they could reproduce they use non deep learning processes and compared them to those outcomes and out of six of those seven and six out of the seven cases they actually had similar outcomes with other procedures so only one paper had a significant better outcome with using deep learning that could be reproduced so deep learning is a word that makes a lot of hype so as you can see there's lots of people here there's kind of like the thing oh yeah we have baselines and I just have to give something that's a bit better so I'm not trying to actually make other systems look better because then my system might look worse so there's not actually a lot of research in this special area deep learning but not in the other areas because there could actually be a lot of much research we can actually gain from that and much betterment but we only think of them as baselines and because I have to quite of show that my system is a bit better so baseline might be a coin toss like 50-50 so when I want to show my system is better than that then 50-50 so if my system is worse then it's not better than the baseline but if my new system is better than 50-50 so it's better than the baseline coin toss so maybe it would be better to compare it to something other something better than just coin toss but of course then my system has it harder to be look good so here's some challenges where everyone can kind of work with these are kind of challenges where people can actually take something that works different than research and with deep learning there's like a lot of stuff that works that we can compare to you and but we see that also classic approaches have here quite a huge part of it more than on normal conferences so if I actually want something that works and I don't want to use that much effort and then deep learning is actually not the go-to system okay so the next aspect we would like to talk about is the effects it has on society so what we do know is that we have to look at sustainability so especially the how can I explain something and transparency so if there's a system that does decides on the matters of life of life on that can actually understand the system as a human so there is a system that scores people how that scores criminals are they going to commit criminal offenses again and judges are actually using the system to decide about the penalty that people get so if we look at the skin color of people who get this course and we see that we have a huge difference between white and black people so in the upper left we say higher and lower scores are equally distributed while we see with people with white skin or people who are perceived as being white we see that the lower scores are much much higher so and we have seen that this is actually not true in reality so that we have for the for the same crime we have different different penalties and judges and so it's not fair so people who work in this area have looked at them and say okay so it should have been actually a reverse distribution reverse to what we see here and so we see matters decided that matters of life and that this is something here no one can understand how the score comes about so the company says well out the background or the color of the skin of those people he has not entered this data but in USA it's correlated to so much to to income and to all other factors so it's not even a significant fact whether they have looked at this color a color skin there so the other thing is what is what is a unique suggestion how to handle things so people call something an algorithm but we have to have a critical few on those systems that our decision to do that make decisions for us so is deep learning a thing that we should call something that is a decision-making system because it's a lot of there's a lot of random data there's a lot of statistics and there and it's more like a machine gut feeling that makes those decisions and is it something we would actually want to rely upon as a society so what we actually see is a huge generalizations we take data points from the past that we know about and then we apply them we train them and then we hope that actually if we generalize them enough and we try as hard as we can with the system and the system has to produce a result and it will produce a result whether it sees a good reason for that result or not they try to find a pattern and then produce a result and that's what we have and so that's what people say well the artificial intelligence prediction something or think about something it's nothing it's just learning from past data it's a generalization and spitting this out as a result so if we talk about prediction there is we don't think about the future we are thinking about the past and so this is a question whether we can actually project things whether we can make predictions from the data we have another problem is that people trust computing systems well probably not all the people in this room which is a nice thing to have but in society it's a very prevalent thing AI is great AI is going to save us AI can't do things that we cannot do and we have some examples here so there is this giant initiative everything everyone has to do AI if there's AI in the application then it's I get the funding if I even write AI on my cosmetics I can sell them better it's done by AI so even if I present my company as AI savvy it might even be good hiring people pretending to be computers because computers cannot do this today and doing a restaurant reservation or something like this so we can say well our own AI systems are so great and so powerful and because it's not a human it's a computer which is much wiser and then you can even gain a get a sort of business advantage a very disconcerting thing we have here hey to answer the question yes that's real that's an actual real thing and I hope I don't have to explain why this is really really worrying so what's happening with people from a social perspective if we interact with machines as if there were people as if they had feelings as if they imitate patterns that we have in relationships what happens to us what is going to happen how much bias that we don't have are we going to accept to get to another topic and I hope we only have to get into this for a short time because I have no no answer to this question I don't need to explain to you why avoiding data collection is a good thing but with deep learning the trouble is we need lots and lots of data and this is obviously in big conflict with our interest to not and other people have probably you probably have mostly seen this this chatbot on Twitter which learned within a day to give extremely racist statements so we might not want to have artificial intelligence actually imitate people and another thing we need to discuss which is relevant for us all where we're all working in on systems which are supposed to move something in the world who is responsible a very typical example for this is autonomous cars who is responsible if an accident happens but this is true for so many other systems there's so many positions which might be responsible for this there's people who marketed this there's people who programmed this there's people who talked about the regular device the regulations there's the insurances and then people who might have had another car with another autonomous system who is the responsible one in this chain of responsibilities so who so ever might be the one who might be liable for this this might give you another chain of responsibilities and insurances on how you're going about this and there's no satisfying answer to this yet there has been a survey among us adults said the majority find it unacceptable in some sectors so for one thing to for criminal risk assessment or maybe the resume screening but it's quite unfortunate that this is actually stuff that's already happening and it's a growing sector as a third topic let's discuss the end yeah this is the third point that's usually carried with negligence we said that about all positions but this is true about the environmental impacts of deep learning as well so this is a topic where we have to talk which we have to talk about even though it might not concern us on the first view so you are aware about cryptocurrencies and how they have a high power usage so a typical Bitcoin transaction uses as much current as 500,000 visa transactions do so this is the power consumption for refrigerator for eight years for one transaction in Bitcoin we also have the general problem deep learning needs lots of data everywhere we need lots of data we need to transport the status somehow and this globally increases the number of service centers and this yields to something like 200 to 500 billion kilowatt hours per year it's not as easy to exactly quantify this apparently but probably if you were to think of these server farms as countries then there's only five countries which actually use more power than these two if we look much on a much smaller scale into the training process of individual models then there's actually quite a whole quite a high power consumption which is actually scary it's not linearly increasing it's actually also scaling probably exponentially and then you see the big state-of-the-art systems usually coming from Google and Facebook and other research institutes and the companies then they use energy for hundreds and thousands sometimes even millions of euros because they're using GPUs and TPUs which are hard to get which are expensive to get so by energy consumption and with these chips and with only few companies doing all this and only being able to do all this to be a state-of-the-art so this has societal impact this obviously has the environmental impact and the power consumption ended the trend in the absolutely wrong direction if you're looking into the co2 equivalents of this then we see that the training of one model which is published this uses as much co2 as five cars including their production and all their fuel fuel during their lifetime and this is still happening for publishing because people want to publish so on a grand scale we're still publishing a new publication with just a minute increase with this carbon and now I'm giving you the happy news yeah just joke but still it's about Google they tried to use machine learning to optimize their data centers get get it to decrease its energy consumption with reinforcement learning in case you know what that is and they dropped their energy consumption by about 40 percent that's good news and yes I'm aware of the irony that we're talking about decrease in energy consumption of a data center which might not exist if we didn't have this discipline but still we might we can use this for our advantage so this might not be applicable to all energy and carbon relevant industries this might be very problematic for car production processes we cannot just exchange with task is done when so when the factory is not in use or so this might not be workable but it is something we need to think about with energy and CO2 that's not the end of the line of this discussion it's also about infrastructure and how do we use that for building for transport for network it's about what kind of space that we need to build those server farms and for the projection of CPU and about working force and resource research resource and our world's resources like metal and oil and there's so much more resources we need for this and it's not just energy and awareness on this point is way too low to actually make good statements about this and so here we are at the questions how actually can we go on so important it's really important that we all are responsible the people who do the research who built the systems we are the ones in charge of building those systems and that those systems I actually built and we are responsible to make them easier and more generalized and if we build systems into cars that use 25 gigabytes of data transfer or something like that and we built stuff for industry and and then there's usage for key in skincare products and of course you can do that and you can earn a lot of money and those in this sector and but it's not a good idea and everyone should actually think about what are the consequences and what do we need to change to actually further this field and it's the nice thing in research it's community and everything is driven by the community it's every researcher can decide what they want to do if they want to go on in this way or if they want to put more effort into those things especially for the society we have to have discourse we have to talk with broad part of society about what do we want what do we think is acceptable and what not which decisions have to be taken it's just not a decision of five people but of the whole society and it's not an easy thing and there's not a clear answer so we have to talk about this we have to further this discussion on all layer like on all fields on all parts of society I have to talk to users a user who should know what it means to use a system what are the consequences of using a system and everyone who uses those systems and the policymakers they have to know about the systems to actually can know to be able to know what reasonable and what not so we have to talk about how are they work where our models are from about reproducibility about responsible ability so in the end we have to rethink instead of just follow so we have to think about these things quite foundationally we're not in a devil spiral we can use the alternatives for deep learning we can actually work with less and we can use this resources so our biggest responsibility is now to talk to people to work against the knowledge lag in the public eye and thank you for your attention and we hope we gave you some new parts and now we have some time for questions but we are really glad if you inform yourself if you do some research if you come to us with some questions thank you so while the plaza is running thanks for listening to the English translation of the deep learning hype your translators were Florian and please give us feedback email us at the hello at 3 3 lingo.org or use hashtag C 3 t on Twitter and now for the Q&A okay so first question AI for cars is very fascinating it decides between a tree and traffic sign I'm really disappointed if I look at AI for SEO search engine optimization well my question is or what is the problem I think is the data so if you look at a tree it's a tree but if there's a website what's the best website or what's the best video that's really a matter of taste so what I would like to explain is wouldn't it be much more useful or required to think about how training data how what's the qualitative level or what's the qualitative ranking of training data yes it's really important to have qualitative high quality data but another problem is it's not really trivial it's there's so of course their search engines are not quite trivial because what I am looking for might not be the same thing as another person is looking for and of course there's like huge amounts of data in the web there's quite a lot different and quite more data than was autonomous cars but especially like with so search engines we have to talk about the data sets and the quality so another question from the internet should we leave the deep learning should we let it go or is there's like some application that might be useful well just letting it go that's probably not the right thing to do it's useful for a couple of things so there's something where it works really well especially if you look at very very complex things there's very little approaches that work as good for example language processing has made a huge step forward by deep learning because human language is so complex that all other approaches we have like accounting syllables comparing letters that didn't help I need a lot of knowledge that goes in there and so I had to think about is it is it the right approach for the thing at hand and well I want I don't want to make it now an overall answer you have to think about it case by case yes that should be the message here so we're not asking should we use the learning we are asking what are we using deep learning for so I'm trying to do this somewhat chronologically with the questions in the room so one question about reproducibility I was talking I was sitting in a lightning talk where someone was talking about he could not reproduce such results his main idea was to yeah just force people that they have to publish all this along with the paper is this something where you are seeing that's right thing to do well there are some certifications at conferences that we have there's probably some at journals well it depends on the field of study so in some areas there's not much publishing in journals because it's easier in conference and it's much much quicker because journal publishing take a long time it's much slower we would wish for having more of that but I would say well the elders of research I would call it that organized the conference they have to actively decide that's high on our priority list they have to enforce it and it's optional right now so we would wish for it definitely and well we're also talking about regulations and then there's public funded research and private research and there's that's very very different challenges so let's now talk to microphone number seven hi thanks for the talk are you thinking that general artificial general intelligence is something you can think of yes well currently hell no okay that was not a professional answer but currently we have very very specialized expert systems that can do one detailed task even in language assistance language processing it's very very restricted so we have done some procedural process but you can you can really break the system so give me some system with three voice measures I can break system there's an American professor actually does that and we have very strong limitations in the upcoming years personally I don't see it happening so but it's something we have to keep an eye on I wouldn't say there's no threats in that and it's also not the focus of the focal point right now so people are working on expert system they they're working on some systems that actually decide which expert systems do I use but the research to have a like a world understanding system which can give arbitrary answers I mean there's interest in that but that's nothing that's really reflects in the current publications because we are not that far and the expert systems are much easier to build and if you're interested that it's called semantic modeling because we have to model the common knowledge and model common knowledge that's not there so we have to work on that so let's take another question from the internet I'm supposed to give you kind regards from the 120 you might know better what that is the question is is reproducibility only trouble in deep learning or is that more trouble and machine learning in general yes it covers it's it's a huge problem in almost every scientific publication that's a huge factor there's some who are more vulnerable to this and some are less vulnerable but but it's it's not as widespread as we would like to use it and this affects all of computer science so everything we said here also affects machine learning in general but especially the deep learning because of the huge data amounts that we have very very strong effects of things here and well since it's a buzzword right now that makes it vulnerable at this point microphone number eight so let me connect to this I'm thinking this is very biased to the publication as they they're playing with the data so long until they have a result that's another trend in the psychology where they had that massively and they solved this by registering the publication into the journal before they know the results so they might have a negative result are there such efforts in machine learning so you're giving the corpus of data before you want to publish oh that's a hard question to answer for that area because I think it wouldn't work like that so published your data well there's data conferences that are specialized in this but you can do a lot of things on one corpus of data and so that's not really helpful I think that the question is much more complex here I wouldn't know of any efforts in that direction I mean it would be nice yes it would be nice to have but I think that's not being done right now I have never encountered in the large conferences no one ever forced me to state beforehand what I think would be the result only if I have my results I would I will just publish that there and if I have my my mistakes and errors in the paper no one is enforcing this there are some approaches there like encouraging publishing failures but a lot of people just frown upon this and say oh whoa that's nothing we are we would like to deal with and that has the effect on machine learning that people do not work systematically in machine learning they just go with a gut feeling and there's a typical sentence well it's understood that this doesn't work yeah what do you know it from if you don't if you don't check it so we have time for only one shot question I'd like to know two things about this black box so I know that you can look into the feature maps of this network and to investigate these some of the impression it's not as black well that depends on how how it works so there's this approach of explainable neural nets excellent excellent and that's being worked on but there are architectures that are completely ununderstandable for us and the the approaches to understand what's working there there are these are our approaches they are there but it's but they only check the model in there but the whole pipeline of machine learning is much much longer it starts with getting the data selecting the data processing the data selecting the features doing the post processing doing the evaluation matrix and these are all things you can tinker with and you have to understand these to explain the whole thing and yes there are approaches that work in their fields but that's it yes and that's the end of the talk thanks Natja and Benjamin