Thorsten O. Zander is the Lichtenberg Professor of Neuroadaptive Human-Computer Interaction at Brandenburg University of Technology in Cottbus-Senftenberg and the founder of Zander Labs. In this talk delivered at LabWeek Field Building at Edge Esmeralda, he explores the latest advancements in AI tools that aim to operate autonomously and at a human level. Learn about the exciting progress in self-driving cars, AI models like ChatGPT, and how these technologies are paving the way toward a true representation of the human mind in machines. Discover how Zander Labs is pioneering the integration of human brain data to enhance AI learning and understanding, ultimately striving for AI systems that can think and adapt like humans.
Created by Protocol Labs and co-curated by Foresight Institute, LabWeek Field Building gathered leading individuals and teams from frontier science to drive progress. The weeklong conference took place at Edge Esmeralda, a pop-up event city in Healdsburg, CA, from June 10-16, 2024. For more info on LabWeek Field Building, go to https://www.labweek.io/24-fb.
Subscribe to our newsletter: https://bit.ly/PLUpdatesSub
Follow Zander Labs
Website: https://www.zanderlabs.com
X: https://www.x.com/zander_labs_BCI
LinkedIn: https://www.linkedin.com/company/zander-labs
Follow Protocol Labs
Website: https://www.protocol.ai
X: https://www.x.com/protocollabs
LinkedIn: https://www.linkedin.com/company/protocollabs
Follow Edge City
Website: https://www.edgecity.live
X: https://x.com/joinedgecity
Follow Foresight Institute
Website: https://www.foresight.org
X: https://x.com/forsightinst
LinkedIn: https://www.linkedin.com/company/foresight-institute
hey hello yeah uh my name is toson and I’m a professor in uh Germany but also the founder of the company zener laabs and today I will speak about the newest results we have in the company and we’re aiming at something I call new adaptive artificial intelligence second so yeah I think uh tool have made a very important contribution to the development of mankind but usually we use tools to um create something but we use tools and um our vision today is that we have tools that are more on our same level that can do something autonomously that represent the human mind and we have made some great progresses like self-driving cars you see here or of course chat jbt and that’s amazing it’s awesome to see that and that’s a huge Improvement uh of how we approach the way to create artificial intelligence nevertheless it’s still far away from a real representation of our mind in a machine the machine might be able to work in the real world but it’s not able to understand what’s going on in the way we understand what what uh the world means so the question is how can we achieve that how can we create an artificial artificial int artificial intelligence that is as good as our mind is and this is what I wanted to to discuss with you today so we have an approach that is utilizing the human mind to train an AI when we look at uh the human mind evolving over time um let’s start at the beginning we have a baby coming to the world and um equipped with a great brain a lot of sensory input Eyes Ears other sensors and it’s constantly getting a lot of data and this G data is annotated by the parents um providing a lot of uh additional information to understand the world and also by the world itself so when I see a leaf falling down as a first time in my life I understand there’s something like gravity and I think we all agree it takes some time for the system to understand the world better I mean um we constantly learn but we need at least a few decades to really be able to maneuver in the world in a meaningful way we now look at in artificial intelligence how we train it today it has computing power and we provide it with a lot of data but compared to the data we receive in a human it’s really not that l large and more importantly the data is not annotated well it’s very simplified data you know we have binary data whatever but it’s it’s not contextualized this data is um partly in itself contextualized but not really realistically in the world and this is why we see great success in chat GPT or large language models models because in the data itself there’s a coherent structure that represents the world to some degree but more importantly the human needs decades to rise and we expect AIS to become intelligent to work uh in the world in a in in a meaningful way after short ter of training term of training and furthermore we discussed the Lin problem a lot the human compatibility of AIS because we afraid that they might do something that uh is not good for us but I find it really interesting to see that the L problem is not really well defined where lot of this we use that term but what it really actually what we mean with it is often unclear and I go with it uh with an analogy with Stuart Russell maybe some of you of know of you know that the unaligned kitchen robot so let’s assume we are on the year 2050 and we have a robot in our kitchen that is preparing food for us it’s doing it perfectly in a way that we like the food that it’s healthy for us and that it’s sustainable and let’s say we eat meat we decide to eat meat and on Wednesday there’s meat day and it’s trying to creates some food but there’s no protein in the kitchen draw so the fridge is empty and then it’s an AI so it tries to uh find a solution and then the dog comes around okay how should the AI know that we are willing to eat a pig but not a dog I mean how how could it know that of we can tell that to the AI but we want it to inere it by itself and the problem is if you have to tell it to the AI to the robot in this case we have to explain it verbally in a meaningful way and there are so many things we would have to explain if the AI couldn’t uh inere it by itself what is good for us that it shares the same value system that we have and my solution for that problem is to learn from the brain because we have the brain as a system when you look at this at this in system systematic way that evolved in the world that can understand the world and that is um generating information about how we process the world and when we can capture that information and translate it to the machine and embedded into a machine the machine might become something like what I call a dig digital digital copy of ourselves so it is copying our conscious interpretation of the world and that is what I think how we should use our technology to create artificial intelligences so of course we need still uh Transformers and whatever but we should annotate the data we have in the real world with our brain responses and of course for that you need an interface and luckily we are working for almost five decades on the development of um brand computer interfaces you’re aware what that is and you see the closed loop cycle of an of an BCI and classically it’s used for Direct Control so like replacing a button press on a computer um and with that restrictions is mainly used for people with disabilities but I propose to use it as a passive prank interface because the brain is exhibiting information about different mental processes like mental workload emotions interpretations like this is wrong I don’t like what I see here or Contex related responses like I’m surprised to see that int what we what we want to do and when we use that um we enable a machine to adapt to the operator implicitly so it does it without me needing to tell the machine what to do it understands because it’s observing me so it might be even without my awareness that it adapts to myself and with that we can the Mach machine is able to derive our ideas Concepts and intentions from our actions itself and we have shown uh different examples in the last two decades how new adaptive HCR human computer interaction could work but what’s next and I would like to talk about what’s next now so my aim is creating new adaptive AIS and I do that with in my company where we received a funding for a project which is called project nafas new ad new new adaptivity for autonomous systems it’s funded by the German government with € 30 million and we’re trying to overcome the obstacles which are currently there to bring bcis to the real world but also create first examples of new adaptive Ai and I would like to show you how we would do that so the first thing is mobile and secure EG when you look at at EG caps right now you see big caps lots of work to apply them look awfully and what we want to do is we want to go with unobtrusive mod self- applicable EG by placing them behind the ears for example so here you see the first prototype a seet developed by a colleague of mine and we have access to that and we will further develop that so you put that behind your ear and you get some data uh emitted by the brain but we want to extend that by putting other electrodes on the front frontal front where no hair is growing so uh it’s easy to apply as well and it might look even look nice but I think it’s not enough I think we get a lot of good lot of data there but we need some data from the center of the head so we are looking now uh to find a way to place two or three electrodes on the top of your head that you can do it yourself but that’s not it of course we need a small amplifier you see a prototype of that so that’s the same thing as you can buy now as an e amplifier which is at least this size or half of the size and here you see an example of how well the system works with the secret on both sides for workload detection so the gray bars are the standard 64 Channel cap when you use a standard state-of-the-art EG system and the red bars are showing what we get when we use uh both s side of the Year secrets so that’s a loss of 2. 2.5 percentage points and the next step is that we do all the processing of the BCI on a trip we don’t do not upload it to the cloud um you’re fully aware what you’re doing there you can select what kind of information you send out because I think if if people want to use a BCI they need to trust it and um I do not know I I do but many people do not know what can be inferred from the EV signal you send out but if you have a button happiness surprise error you know what kind of information you send out you can uh select to to send it out or not the second problem we have um to overcome is the calibration of the BCI this usually takes up to an hour to calibrate one single BCI and um so we take data usually data from a single person and what we we going to do going to do in the project NAFA we collect data from 4,000 people and calibrate BCI on that and that BCI is not indiv individualized so it doesn’t work only on your head it will is transferable to any other person so right now I have BCI for a detection on my USB stick that would work on your head as good as any individually trained PCI we would train right now and then we want to uh have it applicable in different applications so that it’s not only for certain application but that you can you generalize it over applications as well and for that we have already examples of um workload or error potential classifiers that work like that and we are going to develop here up to 15 bcis uh reflecting 15 different states and the big important step here is you can run them Sim simultaneously so right now you can do one or two BC because it takes up to an hour to calibrate them but there that’s a pluck andplay system you can uh get 15 different mental States at the same time without any effort and that reflects your full mental state not only a part of it okay what are what are we going to do with that something I call new adaptive reinforcement learning so assume you have an AI I chose chp’s logo or open AI logo but let’s assume you have an AI and that has a model about the world the gray bars are the model of the world and then we have a human in the same context the the human is perceiving the same thing as the AI and it sees the action of the AI then it can of course it will the per person will interpret the actions of the AI and it will tell whether they like it whether they think it’s good whe they’re happy about it so you get a multi dimensional response on that reflected by this Vector which is fed back to the AI that it can adapt its model it’s like a multimodal reward function and furthermore you can explain the meaning of these Dimensions to the AI you could say X1 is the level of surprise X2 is the degree of error X3 is the current workload and when you take all that information into account the AI can understand in a better way what it’s doing because you give meaning to the values you have there okay but that’s thetically theoretically but I have a proof of concept for that it’s just one-dimensional in this case so we invited people to our Labs equipped from the 64 Channel EG so that’s was in 2016 and we let them observe this scenario here we told them there’s a cursor the red circle with a white dot inside and that is going to jump to adjacent notes but the target is it should uh reach the Target in the right uh bottom corner and the first run was just the cursor was running jumping randomly and they just observed it so that they would probably internally say that was a bad move that’s a good move another good move but it’s totally randomly it will not jump to the Target here very unlikely bad move they were just sitting there observing that Target they didn’t have specific task just observe that and internally judge what you see there of course after a while after 8 minutes or something it would reach the target accidentally but we learned something from that we didn’t tell that to the people but we learned from the brain signals we understood how the brain signal looks like when you’re happy with it and when you’re unhappy with it and we used it for reinforcement learning so we Ed that the the brain signal as a reward function for um enforcement learning so the length of the arrows indicates the likelihood that it goes the cursor would go and let’s assume it goes in the first step to the right hand side one step then we could ask the brain and it would provide this signal for example and we already learned that is a good signal so we would increase the likelihood into that direction and for the next jump and then the next jump happens and let’s say it goes to the upper left corner which is probably a bad move and we get a different response in the brain and we know already that’s a bad move so we reduce the probabilities into that direction for the next jump already after three jumps the cursor knows oh I should go to the right or maybe the upper right and a different perspective on the same thing is that the cursor is building up a model of your directional pref preferences it’s building up a model of a part of your cognition so that’s not a full cognitive copy but it understands what you like and what what your values are in this system so the cursor would jump somewhere you get a brain response and it would adapt its model it has had before and would Contin Contin conly do that until it understands what you want to do so the next video will show you how that looked like on the left hand side you will see the cursor moving around on the right hand side you will see the model the people this is real data but the sub the participant only saw the left hand side only they didn’t saw the model didn’t see the model so first uh jump was bad we reduce the probabilities the model is adapting step by step and you see there’s a coherent idea this direction is the best Direction and the cursor is reaching the target very much faster we also applied that to a new set of grid which is larger and also here the model takes a little bit longer to to to take form but step by step the cursor learns what is a good way to go for and we ask people do you see a difference here to the first run and they said yes it’s going much faster we asked them do you know why and they had ideas like the target is more often in the upper right corner or something like that but nobody of the 24 people had the idea that they were influencing the target of the cursor so also here on the 6×6 grid does some some mistakes but it reaches the target very quickly and without the awareness of the person and when we look at the statistics you see that on the 4×4 grid it takes roughly um on median 27 27 drums to reach the target accidentally and N drums on the 6×6 grid with our reinforce it only takes 13 steps on the 4×4 grid and 23 steps on the 6×6 grid that’s uh a huge Improvement of course but how good is it so we used um perfect reinforcement so we told the cursor basically where the target is and reinforce its actions and you see that would need 10 drums with a perfect reinforcement and 14 gums here so we’re really close to that with that simple form of reinforcement learning with the onedimensional BCR put we could make the cursor reach the Target and here you see a trajectory how the cursor moved it started here moved there and here you see the model how it uh evolved over time and you see this is the result of that not everything looked as nice so there were also some misalignment but in the end it got the right um so uh model and when you look at the signal itself we of course know there’s a good response and a negative response but in this Paradigm we can operationalize or interpret uh the each move differently we could say we could color code the degree of deviation from the target so black would mean it goes directly to the Target red would mean it was goes to the opposite and in between we have color coding in different angles and when you look at the Erp so here’s the jump um here’s the amplitude of the EG signal um and here’s the response you see there’s a linear uh dependency of the degree of error with the um with the brain response so you not can you can not only see good and bad but also any fade in between you localize that with a quite complex method in the prefrontal cortex and we could relate it to predictive coding so we are tapping into the predictive coding of the human mind which is basically the re the reward function of the human mind this is how there’s a theory out there which tells us this is how we learn ourselves so we can tap into higher order of cognition and we’re assessing the reort function of the brain so that is proof that this reinforcement learning works the last example is a little bit more complex when we think about how is our our intelligence that rises in ourselves we know that biological learning is categorical what does it mean so one way to uh train a biological system is clicker training you know you might uh have heard about that that you can clicker your uh reward your dog when you click on good occasions and that is not the optimal learning way better one is that you have categories so I have an idea what a um cup is and I could pour water in it and I’ve learned that in my life and for the first time I see a mug and I see the mug is similar to the cup but it’s not the same thing so I transfer the knowledge I have from the cup to The Mug but have a new entity and adapt that independently of the other one and then I have subcategories and whatever and whatever and that is a way better Theory to describe human learning than clicker training finnally there was a TV show in Germany recently clicker train your baby where they proposed that you don’t have to teach anything to your baby you can just click or train it it’s it’s really good it came from the UK so that was nice TV show okay but how when we look at um how do machines learn currently so how do we do machine learning we all speak about machine learning how we do that we have um supervised learning two classes good and bad or we have reinforcement learning but that’s nothing else but clicker training it’s the same thing so why if we know clicker training is not the best way to do it why do we do it all the time why don’t we use uh categorical learning on on the machine the problem is we can’t do it there’s no way to implement the categories so my mind has probably millions of categories in itself so if I would be able to access that I would need a long time to transfer that into a machine if I would have a machine that could explain what my are it would already be intelligent so that’s a real problem there’s no way at the moment to implement categories in an in an AI there is theory about that what would happen if we had that and it’s showing tremendous increase in reliability but we don’t know how to do that yet but I will show you that we can do that with the passive PC as well so here we have a task it’s just a proof of concept again you see letters q DB and P which are the same letters you can rotate or flip but it’s the same symbol and then you have a square or circle around it and we invited people to interpret this um selection of symbols and we gave them a random uh rule let’s say the question was are there more circles in a b then Q is in a square and then they have to judge yeah it’s true or it’s not true that’s one example if you can exchange these rules by any other rule it’s just a random Rule and our theory was that they would identify firstly the bees in a square Q’s in a squares and the bees in circle as relevant and they would also differentiate the BS in a circle with a different brain response compared to the q’s in a square so the brain would emit a different response and we could collect that with the Ed and the response to all the other symbols are is a different one so we have three different three categories of brain responses the red circle the green circle and the yellow ccle circle and they represent the different types of similar we see there okay but of course this is not is solvable problem so you can use standard machine learning to solve this problem right and when we do that we have this approach here it’s a state-of-the-art way to uh solve this problem and we get this result you know after 1,500 trials we have 82% of accuracy if we do Super Wise learning and here you see the trajectory it starts at 50% goes a little bit down and then up and it’s line are learning this is what we expect from from State of-the-art machine learning and now it took the same model but inserted the brain response so we gave an additional sensor add additional information and here you see how well that worked the blue curve is showing the result on the same training but with brain signals so you see at the end after 1,300 trials it’s already with 100% accurate it doesn’t do any error after that anymore and in the beginning after 150 trials you’re already at 68% it learns very quickly and continuously with 20% of improvement and that’s huge usually when you have an improvement in accuracies of AIS you might have one or two% points and you’re happy and we have here 20% and that shows that passive pcis can also improve the learning of an AI in a categorical way and I think if you do it more extensively what which we will do in the project we can create a lot of categories in an AI and we’ll understand the world better as we do so here again uh here you see the things we do in nafas the project we will first develop tools like the EG headset um the classifiers databases so we will collect 4,000 data sets which we will make available to the public so you can use them these 4,000 data sets in probably 3 years and then we will have uh an implementation of the new adaptive HCI new adaptive AI part and have several demonstrators in there okay concluding in the last two minutes what I think what we should have is a convergence of the human mind with artificial intelligence so when when when we want to make this robot more capable we need to integrate its own processing with information about the human mind the human mind can annotate the world and make it accessible to the robot and give meaning to what it sees there currently the robot might name a I can describe what it sees here there’s a cup there’s people sitting there and the camera and whatever but it doesn’t have an idea what it means and when we annotate it with the brain responses categorical learning reinforcement learning I am sure the machine can get better and it might become an equal partner to us which is totally aligned with us what do I mean with alignment with each person so I don’t think that we have alignment in a group but we might have an AI your personal system that is totally aligned to yourself in that way that’s it and we’re hiring