EthCC 2: Kazuaki Ishiguro

I give you all right thanks everyone my name is Kazuaki I'm from a company called Kruger based in Tokyo Japan and I'm also working on the protocol connector and then I will talk about how blockchain ensure the reliability of data in AI so today I will talk about what we do first I know sir I will quick review of AI history and its composition right now and also I would like to point out the problem of AI industry right now and also of all share regimes and solutions regarding to reliability of AI its dataset so it's it's called connect oh it's a official science human science term of the connectome is a comprehensive map or neuro connection of the brain so basically it's a diagram of a nervous system based on Wikipedia and also our project is called connectome basically our vision is towards the world where human beings can trust AI and also we provide some technology and human beings so we provide some AI assistant with uh with a human look human-like visuals so uh I'll show you some demo I already tell looks like we have a visitor yes I am yes we do nice to meet you I am Rachel hi Rachel where are you from I'm from Germany you came from far away that's true have you ever been to Germany no I have not but I will be going there in May if you have any questions about me please ask how old are you that is a secret where are you from I'm from Shibuya Japan what do you do Rachel I am working at couger as assistant I can provide some assistance for you what's your hobby I have started to learn dancing recently Oh can you dance for me okay I will try I don't want to I want to talk about the definition of dancing but that's that's that's her definition so that's that's the Rachel that we are developing right now and they all talk about how how it will relate to the security of data in AI so basically AI is kinda like a training model so that's an AI and then if if you want to use a AI in some systems but basically you have to have you have to give some input today I and then you get output as a result so input is for the Rachel side input is also called comprehension AI and the Rachel has a vision which does object detection in the face recognition remember you know the richest are saying hey hey absolutely or hey hey someone that's a face recognition feature and no sir scene recognition either the person in the sitting or dancing or or walking or whatever and also of course as you know the Rachel has a sound voice recognition so that's that's the input version and also as an output they do evaluation if if a Cherno the person right in front of her or not and also another version of output action is a character AI also which is called called game AI on the game game industry so the game AI is a little bit different from a current AI system that you are using because the game game AI you have to change dynamically based on the output and then based on the output the decision-making process will be changing on the real-time so on our system we provide some scenario data which is just like simple script version Oh based on now the reaction of our output so I would like to point out AI problem right now it's it's if you are engaged in AI you already know that's AI situation is currently in the black box because in order to create some train model you have to help you are required to have tons of data sets and no so you are required to have a resources which stores data or or the person who put the label who give you a correct answers and those are that will cause the data silo and then they they are store in the block books because you don't want to share to somebody else you don't want to share it to competitors because the data or correct data is so much valuable today and now of course those you know tech giants are doing it Gugu Amazon Facebook Apple louver or Alibaba in China so I would say AI today is not trustworthy your you might already know smart speakers like Amazon Alexa or Google Holmes has a bug and no so you know you never know those speakers are currently monitoring your conversation or not how can you guys tell and has a counterpart of those block books they are currently xai or explainable AI type is is a trend right now Xai is basically the idea is the based on the output coming from my China model you have to be he has to be explainable why your you AI made this kind of decision but but there's also trade-off of explainable AI because in order to make explainable AI the AI or train Morrow has to be real has to be has to do only like very simple tasks vice versa if you want to do some like complex detection on the AI side it's really hard it's make it's really hard to make that a trained model as explainable so in order to tackle those problems we develop one like the project called gene flow that was proposed in AU Kong trono last year so this is a very simple overview like three entities the the basic idea here is we would like to store the data sets and then script and thereafter the epoch training epoch in the way we got trained model we we create as a package in the store in the ethereum blockchain so that we can always like trace the gross history over AI so when we are usually like when once we you got trained model you have to keep updating with a new data sets right so once we get like new data so we update the script and then we create another transaction and going on and on like this so you know once a I got rogue or AI got bug or something you can always trust back once when you're trained model was working correctly so that's the basic idea and then there are lots of AI in the blockchain projects that they are working on this kind of project more like security or providing a security environment so I really recommend you to check that out so this I would say you know as and one option the complication or input AI is can be safe using those kind of technologies so what about those two decision-making processes interaction reaction process so the simple idea of decision-making process is something like this if my AI o chão and model have received some input and detects something to do something so that's a basic idea for example the object detection if you get some some input has a pictures and if you detect some person you know give a reaction as a person so the very simple version and there but from our point of view is quite different because you only get the output form III right and there like people think you know there's some magic happening inside of your AI so like maybe your your once you're a I got wrong you never know and you know we have to we have to you know because because my algorithms they sell so you we never know so that we would like to tackle these problem as well and that we said another proof of concept which is called smarter space use cases the basic idea here is like we we put the sensor and they capture the image every seconds and then encoded into base64 nests in the crate and use transaction and you know at the same time we run those sensor software on the Raspberry Pi and then we running a object detection as well so that's that's kind of a simple architecture that's more space use keys so we we we capture the image and then you know encoding as metadata and they're running on the object detection we use the opencv version with the SSD mobile net which is also like open source data sets and once we get and then the important part is only two here so if my images capture more than three people people in the room with city but if those people don't have a hot water bottle like this that the virtual human agent recommend some you know recommend to recommend someone to bring that bottle to the different rooms so that's a basic idea and then I have a demo here so all share to you [Music] kochira the number of people have been increased in the conference room they may stay long because they have seen it why don't you take some drinks for them you need two of them because we have two more people sure we're two Bonavena yes please bring two bottles all right thank you kochira [Music] okay so that's the basic idea and there we can we can you know apply to smart home or or home security system so that's the basic idea of that PR pre-hook concept so we can say there's a decision making process in an output is can be saved using the blocking and mother AI developers in also a game developer - so we are developing a connectome SDK so that all the developer can create your virtual human or your new AI models on the game AI version we we provide some desktop application running on the client side which is the simple version of election application and that we provide GUI so you can actually you don't have to you don't even have to write some codes you just need to you know dragon job and set up some parameters how you're a virtual human want to want to act and then it will be running on the PC or mobile or tablet and also uh as I mentioned earlier we will like to we are thinking about integrating the gene flow version so that you can always trace back what kind of you know data or or train model will apply to this the VHA currently you are running so you can always you know get the hash link and then you can with trace back on the first process you know we are going we are thinking plugging to other providers such as like data provider or computation provider for for more secure environment computations and also we we are very welcome to work with some 3d model designer if you want to you know if you want your virtual agent want to help different looks so we are going to help provide some marketplace for virtual human so let's say you have you got new virtual agent on your own but you know currently lots of different virtual human providers such as magic leap or some other Japanese creators work Jana and then we I really hope you know those those entity are selling their own virtue Asian or Noronha Network but as we growing up the number of virtue agent it will be really hard to find your own favorite virtue agent on the market so how can you how can you evaluate how can you find your favorite virtue agent or how can you tell there might be some like evil or some you know malicious virtue agent or some you know virus in it how can you tell so how how to evaluate virtue agent with our diversity as we think you know think about the web 2.0 version for evaluation or search engine mechanism it's there are tons of problems like sponsored content has a priority or or I really hate that s CEO and also the filtering algorithm can be in the black book so we never know and know so you know there's a untrusted feedback so people can't money plate some you know good or bad feedback so we we are going to do on the Descent role as weight of course as a crypto economic fashion we are going to you know rely on the community our we want or our marketplace maintained by community using like incentive mechanism or game game theory we are going to implement token cure registry or also a bonding curve mechanism for for price mechanism for example we have done TCR together with Ocean protocol from Berlin and then we we set the sharing points like what is the best you know customer service virtual human nation and as for the boning curb press mechanism we have deployed on the loan Marketplace on the github and then you know when creator deploy some virtual agent and then and then set up the parameter at the same time the price will never be manipulated by anyone else because the price is price mechanism or price formula is store in the blockchain I mean small contract so I will probably we are providing some interface for price mechanism it's a very popular one some other bonding curve project is using so if you want your aren't bonding curved formula you can create your own using those interfaces and also we are really focused we focused on the community development last year and we be to North America and we did a some workshop with the Stanford University last year and we did a bunch of meet up there and that we did also workshop in Africa Rwanda and I was so we tours in Japan and China and Singapore and collaborating with our bunch of different project not only interior and if not bubbly not last not least we are going to help the meetup tonight so if you are interested in join us please scan that QR cord and I think that that is full right now just I just check that out like a couple minutes ago but you can still be on that waiting list so please let me know I fir interested and that's it for me thank you for listening [Applause]