[ML News] AI-generated patent approved | Germany gets an analog to OpenAI | ML cheats video games

an ai is now officially listed as the inventor in a patent alef alpha raises 27 million dollars to build europe's open ai and an open source replication of dali is released welcome to ml news all right before we get into all the stuff this video is sponsored by weight and biases weights and biases is a one-stop shop for machine learning researchers to track their experiments save their models recreate their old experiments uh share work with others and generally analyze their results weights and biases allows you with one single line of code to track your experiments which means that weights and biases will track the execution run of your experiment it will track the results it will track saved models and checkpoints upload it all to a convenient central place in your profile and that allows you to analyze visualize all of your experiments and data think of it like effortless tensorboard in the cloud weights and biases has integrations across all of the deep learning frameworks pytorch tensorflow hogging phase you name it they probably have an integration available today i want to tell you about a new feature that they have which is called tables now the name is deceptively simple uh table is simply a grid of stuff but in weights and bias these tables allow you to view things like data sets but also outputs of your runs any kind of artifact you have you can analyze in tables tables allow you to sort group filter and do anything with the data you're looking at and you can take advantage of all the visualization capabilities that you're used to from weights and biases dashboards for example here we automatically visualize the results of pixel level annotations i mean look at that left hand side that model sucks look at the bottom why is the sky labeled as trees clearly you have to do something here so as you can see you can analyze the output of your runs you can see where the model still makes mistakes by filtering for the samples that are classified incorrectly if for some reason weights and biases doesn't have a visualization for your type of data which is unlikely if they don't have it they allow you to actually integrate with their framework in order to produce one the capabilities here are really endless here you can see we visualize anything from sound files to training plots to spectrograms whatever you can think of so as a special bonus viewers of this channel only get 80 off today of the basic plan which you don't need actually because it's free yes it's completely free there's really nothing stopping you from going there and making an account personal accounts free unlimited experiments if you're a bit more involved if you want a team and if that team is large and does a lot of tracking you'll have to give them some money but their main income comes from big enterprises that want to use this internally if you are such a big enterprise don't hesitate to give them a call and give them a lot of money in that way you'll be supporting all the free accounts for all us plebs there are special options for academic research teams which do get free team accounts and you can also self-host if you need to be compliant with some sort of regulations so again go over to weights and biases and check it out there's a lot of features that i haven't even talked about yet such as hyper parameter optimization that's done automatically check it out and now let's get into the news [Music] i'm back yay what did i miss what has been going on how do i do how do i do news i forgot all right the global legal post rights south africa issues world's first patent listing ai as inventor so this person right here is professor ryan abbott he and his legal team have been fighting around the world applying for patents that list the ai named davos as the inventor of two particular inventions so now they finally succeeded in south africa and also as abc news writes an australian court has equally ruled that ai can be listed as an inventor on a patent application now the situation is a little bit complex and i'm not a lawyer so don't take my word for it but the ownership of the patent rests with the creator of davis of the ai while davos is listed as the inventor so here's one of the things that dabbas apparently invented it's kind of a fractal thing so they're saying this is kind of a food container or something and the fractality somehow makes it good and you can connect containers together but there's also this uh light emitting thing that has kind of a fractalish pulse or something that makes it really noticeable and this here is stephen taller who is the inventor of davos and therefore the owner of the patent now i was immensely interested into this and i have spent way too much time researching this here is kind of a few takeaways first i thought this is a pr stunt come on you know why can't you just list yourself as an inventor because ultimately ai is like a tool right and how does an ai even come up with new ideas like what counts as new ideas and like how does an ai come up with this or or this like what was the part that the ai did what was the starting point what was it do it like i'm so confused okay so this is the website of the team of the legal professionals that got the patents through to through the courts and they answer some of these questions and their claim here is that in the various legal systems the granting of a patent requires the inventor to perform like the invention step like there's a specific step in the conception of an of an idea that is like the innovative step and it is actually criminal offense to list the wrong individual as an inventor so the inventor does the creative step and you have to list that person as the inventor otherwise it's criminal offense now the question is if legally the ai did that inventive step whatever that means technically you should list the ai there because you can't list any of your your employees you can't list yourself because you've only controlled and built the ai but the ai did the actual step that the law requires to be listed under the inventor and apparently they claim at places patent applications have been rejected because of this so from this perspective it kind of makes sense that you should be able to list the ai as the inventor now counter to that some legal systems also reject this notion saying only a natural person can be an inventor and therefore on some of these inventions simply no patent can be granted which would be discouraging from researching stuff remember ai is used to make inventions in such field as drug discovery where the ai simply comes up with new compounds and then you test them so in a way the inventive step is performed by the ai if you could not apply for a patent in that that would discourage research in these directions all right so this seemed to me like to be a reasonable explanation but that's only the surface right here i was much more interested in the question of how how does this system that i have never heard of come up with new invention and here on this hideous website of this legal team this question appears to be answered and cut uh so this has gotten so long through the edits that it just completely blows the format of ml news so what we're gonna do is we're gonna cut the rest of this into its own video because this is really weird this davos system is weird this whole case is weird the too long didn't read is there might be a valid legal reason why ai needs to be listed as an inventor on a patent also at the same time this is probably a giant pr stunt and the inventions themselves aren't they're nothing so uh you know look forward to the next video make up your own mind let's go on with the news all right german startup olive alpha raises 27 million us dollar series a round to build europe's open ai from techcrunch this is jonas andrulis the founder of aleph alpha with headquarters in heidelberg in germany which is not too far from here and the goal is to build the equivalent of open ai but in a european fashion so it says the german ai startup out of alpha has now raised 23 million euro which is 27 million in real money in a series a founding co-led by early bird vc lake star and ubc partners the team says it will have a strong commitment to open source communities such as luther ai academic partnerships and will be pushing european values and ethical standards it says supporting fairer access to modern ai research aimed at counteracting the ongoing de-democratization monopolization and loss of control or transparency so while these are laudable goals and i really hope they achieve and stick to these goals remember that openai has said the same at the beginning and now openai is mostly interested in closing down access to their stuff and charging for it but luckily venture capitalists which are the main founders of this venture right here are not known to ever wanting their money back or anything like this so this should just be a breeze for olive alpha so i wish jonas and co-founder samwell and anyone part of aleph alpha all the best and big success in their endeavors it's going to be fun having sort of a counter force to the us here in europe [Music] robotics 24 7 says amp robotics marks milestone in data pick rates for automated recycling so speaking of companies and raising money this company is now raising series b for about 55 million us dollars and they're in the space of garbage sorting and disposal and recycling so they've developed these analysis and gripper technologies and this is incredibly cool to watch i mean we're always talking about ai taking away our jobs i don't think people will be too sad that ai is going to take away their jobs in this particular field so here the ai automatically analyzes the streams of garbage and sorts them by the materials in them and sorry these blocks of cans just look really cool also there is such a thing as waste expo didn't know excellent must be a blast next news deepmind releases a paper called open-ended learning leads to generally capable agents so what they do is they build an environment called x-land this is kind of a 3d environment and the agents in here you can see on the top left and top right this is what they see apparently and they have to fulfill various goals in these environments you can build any kind of environment you want in x-land then you can tell the agents to achieve that apparently the paper is about when you instruct the agents to learn multiple goals many goals at the same time or after one another they become generally capable as opposed to just having a single objective and then ending up with a very narrow skilled agent now ex-land can be used to not only have many different environment spatially but also have many different tasks or games in this environment so they have capture the flag king of the hill and so on in the paper they actually detail how they use population based methods in order to train these agents how good they are at zero shop learning and so on and this is all pretty cool however these things and results aren't that new we already knew that population based training is probably good if you want to achieve some generally skilled agents we already knew that multi-objective or objective conditioned learning is probably a good thing ultimately the agents here are simply an observation encoder into an lstm and then they take in the goal conditioning and then it's a standard actor critic reinforcement learning i guess what i want to say is that the research isn't necessarily super new or exciting but you can get a lot lot of publicity if you build something that's 3d and looks really cool so if you want you can build your own stuff in ex-land if you work at deepmind because i don't think it's open source so haha the new york times writes something bothering you tell it to robot and it is about the system that delivers cognitive behavioral therapy through an app so cognitive behavioral therapy is one of the more successful approaches to treat things like depression or anxieties it is rather formulaic as this article describes and therefore it lends itself at least a little bit to be incorporated into some kind of algorithm so the article is a discussion of is this good is this bad the pros are that usually a human therapist is very expensive and there aren't enough of them especially in times of a global health crisis on the other hand critics argue that these algorithms aren't yet good enough to replace a human because they cannot intrinsically understand the things that the humans say and you get the idea the new york times accompanies this person right here eli who has tried out the app for a given period of time eli details how the app sometimes fails responding to my boss doesn't appreciate the work i do and i can't seem to get her approval the bot answers with that sounds difficult does this happen more in the morning or at night it is a little bit of an improvement i guess over something like eliza however it still seems to be rather formulaic so my own personal opinion is this if i have some problems there are books that i can read self-help books that guide me through the process of somehow solving my own problems these books are necessarily impersonal they are written by a person but they're not personalized to me in any way it's the same text for every single person that buys the book so if a book like this can help me then certainly a little bit of an algorithmized version of a book like this might help me too you know there are ways to make it worse but i don't think much so if you think that there are good books that have helped you in the past to overcome personal issues or problems or any kind of improvement then it's entirely possible that an app like this does the same thing i don't think we have to necessarily seek to replace therapists but there are a lot of people who cannot afford therapists or don't have one close by and in this case such an app can probably help of course it's also easy to see that people will feel as though that actually replaces a competent therapist and not seek the attention of an actual therapist when it's needed so at the end eli breaks up with robot saying he was unimpressed by the bot's advice for beating back loneliness and despair but he is not entirely sorry that he tried it out the mere act of typing out his problems was helpful and through the process he pinpointed what he actually needed to feel better yes so it worked now eli is seeing a human therapist in philadelphia for 110 dollars a session next news synced writes google's wordcraft text editor advances human ai collaborative story writing so the text editor isn't out yet just a paper and a demo video where a human writes something and then clicks on a button and then the machine sort of continues the story this seems to be sort of a gpt three-ish thing with an interface that just helps you select from different continuations and does the prompt engineering in a smart way for you you can even customize the prompt you can ask the model to elaborate on particular parts of the story and then choose from various continuation i think that's pretty cool if it ever will appear online which i'm not sure uh given that it's google but if it ever will appear something like this might lead humans to just come up with new ideas through this thing so pretty cool [Music] next news pc mag rights machine learning is now being used to cheat in multiplayer games so there's apparently this video here that demonstrates that a bot is used for cheating in games now aimbots have been a thing for a while but apparently this thing works in a little bit of a different way and it also works on consoles which for now has been a kind of a difficult thing for aimbots so what you do is you hook up your console to a video capture card feed that into your pc and the pc would actually send commands to your controller so you'd hold the controller but your controls would sort of be overwritten at times by the input of the cheat engine and that makes detecting these cheats rather hard to use now it just says that machine learning is used in order to control this right here you could also imagine this being just kind of a classic aimbot that just recognizes some pixels and then shoots at it but apparently it's machine learning based so you know it's in ml news thanks [Music] next news google releases the open buildings data set which is a data set that across satellite images of africa has annotations of over 516 million buildings this goes along with a paper where they detail the challenges that they had to overcome to do this so you can devise various failure modes right here so all of these pictures for examples are not buildings the top left are water pools top rider rocks then here there are some buildings but the thing in the red square is not a building it's just a bunch of walls the left are containers this is very difficult google has annotated over i think a million images 1.75 million images or sorry google has annotated 1.75 million buildings in 100 000 images by hand and then trained a system on it the paper details how difficult that was how much you have to use augmentation and regularization in order to do that but in the end they've come up with this giant data set that you can now use you can actually explore the data set in this interactive explorer right here so you can switch between this view which is i'm not sure how helpful that is or this view i have discovered so if you zoom in right here i have discovered however that sometimes i feel at least like this piece here is this an actual building it says it's a very high confidence building i'm not sure honestly also this thing here this might be one but it seems like it works pretty well just overall the challenges are also recognizing buildings in both rural areas where they kind of blend into the environment and recognizing buildings in commercial or dense populated areas where you mainly have to separate buildings from each other so pretty cool give the open buildings data set a try if you're interested [Music] next mit technology review writes hundreds of ai tools have been built to catch covet none of them helped yet another article about the shortcomings of machine learning research and the take of this article is somehow you know more effort is needed and criticizing ml research in the meantime i have a bit of a more cynical approach right here like we've known long enough about the publication pressure in ml research and to use a buzzword topic like kovid in order to get a paper published by simply applying whatever your thing is in research whatever your topic is and using it on some kind of code data set in order to get a publication out of it because people think like oh this is you know relevant we need to publish fast now i don't think the main motivation of 99 of this research was actually to develop something that actually works old methods are slapped onto new topics in order to get publications and we will continue to see that in the future as well don't expect any of these things to work in the first place [Music] next news dolly mini is an open source replication effort of open ai's dali so these people have built a version of dali that is much smaller but has first signs of actually working remember dali goes from text to images and you can actually try it out yourself on an online interactive demo on hogging face here's my query for creepy clown and the model does not disappoint it seems like there's still a gap probably a gap in size model size and data set size until this project reaches the level of dali if ever but still it's pretty cool and i love the avocado chair just as much as the dolly one [Music] okay we come to the helpful library section of ml news helpful libraries first helpful library is kind of big news openai releases triton which is a language that allows you to build custom kuda kernels and these cuda kernels are super duper duper fast and you don't have to know low level c plus cuda in order to produce them so there's a blog post and code to go along with it detailing in in very detail what's now possible with triton and apparently openai has made this in such a way that people who have no previous experience with cuda programming are able to produce kernels that are as fast or faster than the kernels that were previously programmed by experienced cuda programmers so if you have something that doesn't have a efficient cuda kernel yet maybe give triton a try next helpful library flammel fast and lightweight auto ml is a library for cost effective hyper parameter optimization so apparently you enter your problem to optimize and your cost and the library will optimize your hyper parameter towards your cost taking into account how much each hyper parameter setting costs to explore so for example if you have something like model size as a hyper parameter it will preferably try the smaller sizes first because they cost less and you can search more before it then scales up that hyper parameter pretty cool give it a try next up for library italian clip remember clip scores images and text together and italian clip is now available particularly it can classify such things as a and oh i'm kidding it's uh it's a cool project check it out if you are italian speaking or building italian speaking products next helpful library deepmind releases melting pot and evaluation suite for multi-agent reinforcement learning now other than excellent this one is actually open it's an environment in deepmind 2d lab and has various scenarios for multi-agent reinforcement learning and this actually looks like you can do some research with it and multi-agent reinforcement learning especially something like cooperative multi-agent reinforcement learning is one of these areas that is still largely unexplored and we don't have super good algorithms for it yet so if you're looking for some research to do this might be a cool topic there's an old helpful library with some news mojoco the 3d simulator that has been used for a long time for doing things like continuous reinforcement learning control problems and so on is now free the product requires a license but they do give out a free license to anyone at least until the 31st of october 2021.

so if the availability of the license has blocked you so far give it a try now also in rl news openai gym has a new maintainer that is going to address the poll requests that are there project has been kinda dead for a long time and the new maintainer makes it clear that there aren't going to be new environments uh major breaking changes environment wrappers anything like this i think they simply want to make the gym usable and up-to-date as it is pretty cool if you're a gym user this should give you some stability and compatibility with current libraries the new maintainer is jk terry thanks for your work so in last news for today the free software foundation calls for white papers on the philosophical and legal questions around copilot apparently they're contacted understandably a lot with regards to copilot and the kind of legal ramifications of copyright and patents in what copilot does if you don't know what co-pilot is watch ml news from a while ago in essence they give you 500 bucks if you publish a paper through them that somehow elaborates on part of these topics so areas of interest are is co-pilot training on public repositories infringing copyright is it fair use how likely is the output of copal to generate actionable claims of violations on gpa licensed works and so on so there are some submission guidelines and i wonder if there's a way i can submit my ml news segment to this where's my 500 bucks richard come on so the criticism of the free software foundation is that copilot is what they call service as a software substitute which is a term they came up with to replace saas software as a service to make it more clear of course richard stallman here writes the basic point is you can have control over a program someone else wrote if it's free but you can never have control over a service someone else runs so never use a service where in principle running a program would do never richard says never ok new.org let's look at that a certificate what kind of certificate is there hmm details it's by let's encrypt g is let's encrypt a program or a service i wonder what's up richard you're perfectly capable of generating ssl certificates using openssl a free program that you can run yet you elect to use a service like let's encrypt well isn't that a jolly alright this was already way too long this was it for this week's ml news please check out weights and biases they're a great system and i'll see you next time bye bye

2021-08-08 15:31

Other news