Science at Work: Changing the game for Artificial Intelligence with neuromorphic computing

good morning everyone uh welcome to argonne national laboratory's science at work webinar changing the game for ai with neuromorphic computing my name is john harvey and i'm a business development executive in our science and technology partnerships and outreach organization uh just a few logistical notes before we get started for a better experience we recommend turning off your vpn just because of bandwidth also your video and audio were muted when you joined the webinar if you have any difficulties related to the technology we recommend logging out and logging back in to to remedy the situation please ask questions through the public chat which is the little bubble at the top because of this is such a short seminar we won't have time to answer those questions live but we will provide email responses to those questions after the webinar ends finally please note that today's webinar is being recorded and will be made available for uh for viewing after the after the webinar at a later date now i'd like to introduce our speaker for today angel yang was jill who is a principal mater principal material scientist at argonne and a fellow at the northwestern argonne institute for science and engineering at northwestern university angel's current research focus is in two different areas materials growth and neuromorphic computing which is what he'll be discussing this morning angel we're looking forward to your remarks today thank you very much for your introduction so i think that for many if not everybody in the audience can recognize the impact that artificial intelligence has in our lives if you look at the ability to process tons of information and answer complex problems artificial intelligence has demonstrated over and over again that it can change the way we we do things one limitation of artificial intelligence right now is the way it operates it relies on lots of data there is a a lot of training going on and then once you have a training system that system can be deployed say in your smartphone or in a chip but once that system is deployed it's static if there are changes or perturbations in the environment that the system needs to respond to it cannot do it unless it falls within the data set that he has been trained for that's very different with respect to how we operate we and in general humans and animals we have a tremendous ability to learn on the fly to react to new information and adapt to changes in the environment and if you if you look at different applications where ai can really make a change for instance a self-driving vehicle technology is one example what you would like is to have a system that is able to recognize that something has changed and with a few examples being able to adapt and recover pretty much like we do that is something that has a renewed a lot of interest in the idea of using brains as an inspiration for computing and that's what it's called neuromorphic computing now traditionally we look at brain inspired computing and we think of us humans and how complex our brain is and the tremendous capabilities that it has what we have done at argonne is to take a step back and look not just at humans but other different types of inspirations that can help us make systems that are essentially capable of doing this type of a smart learning and in particular one of the most promising things that we have found is insects insects for us is not just a an inspiration for making a a compact ai but if you think of the capabilities of an inside it's almost like an a smart sensor that is capable of operating in a noisy environment new information and as it learns it can adapt and it can keep working even if there is something that could otherwise cause a catastrophic failure if you think of bees they can forage miles away from the hive and go back and they can learn to go to specific places depending uh on the the time of the day that is the type of flexibility that we've been very interested in and that we have been exploring and tried to port into hardware so how do we do that how do we go from an insect into essentially something that can go on a chip or can help us understand how to make better chips so we what we have done is to break down the problem in two different parts first of all we take a look at all the fantastic research in neuroscience and instant neuroscience and in behavioral analysis of insects that it's out there and what we try to do is to extract the uh the mathematical principles that help insects perform the way they perform and once we have those mathematical principles we can run them in the same way that you would run a machine learning or ai algorithm and we can compare the performance of something that is inspired with the inset brain with what is essentially stated of the art in artificial intelligence and machine learning and what we have done when we do that is to identify certain characteristics that are very unique to insects that help them perform even though they have just very few neurons with a very small brain one of them is for instance the ability to take one network one connection in the brain and adapt its functionality depending on the context and that is something that helps them be flexible and adapt to different tasks or different changes in real time while reusing all the circuitry that you have there's that's something that it's a very important the second aspect of the insert brain is the ability to not learn everything they come with a pre-trained component and then they're learning only the specific information that is needed to carry out that task and that's again in contrast with the traditional machine learning approaches where you're training the whole network to essentially carry out a very specific task so when we and our collaborators from computer science take these networks and compare them with benchmarks in machine learning and in particular in a domain called continual learning that requires learning progressively new information without forgetting any all information that they have learned it turns out that even though they are very small and very nimble they can perform as well in some of the tasks as the state of the other algorithms that are out there so that's something very promising the second approach is you have this mathematical model you have an understanding of what makes insectic and make them smart and what you want to do is to break them down if you want to pour them into hardware and you have to understand how to best translate these principles into something that you can put into and again the approach that we have taking is that we have targeted three different type of hardware approaches here first of all we have looked together with collaborators such as the university of san antonio to use a off-the-shelf a resources like fpgas that you can program and you can tune to perform the machine learning or the ai that insects do the second thing is to look at state-of-the-art uh neuromorphic chips there are already chips out there like intel's lawyer that are capable of doing some of the things that we have observed in the insects in neuroscience and we can therefore port the algorithms and compare the performance with what you would expect from the mathematical standpoint and the final thing though is that we can take these ideas and figure out what new capabilities how we could change the way we design chips using noble materials and in that sense is where we are bringing together two different areas of expertise because argon has extremely strong program in a technique called atomic layer deposition that's a thin film technique that is used in semiconductor processing so what we are doing is to use a so-called co-design approach in a code design you have an application at the very end for instance we've been able to apply inset networks to a process data from a wireless the type of data that you would have coming in and out from your cell phone and we say we use it as a target to simultaneously design not just the architecture that it would work but what type of novel materials do we need or how to best integrate the materials that we already have into uh into these architectures to optimize the ability to learn in real time and the results that we have obtained it can be applied in not just to silicon but we can move beyond silicon to apply to different types of substrates for computing one example for instance is that we have tried to extend the thermal er the environmental conditions in which you can compute to move computing towards extreme environments and we have found that the combination of these node materials with other non-silicon platforms like silicon carbide it can help you maximize the amount of compute that you can carried out while minimizing the number of components that is needed which is something really important when you move to temperatures as high as 300 to 400 degrees c and even high radiation environments so in summary i hope that you have found interesting this this approach we believe that there is a lot of promise to target these techniques to develop new sensors and chips that are specialized for real life applications and if you are interested in learning more please contact us and thank you very much for being here with us today great thank you angel that was a very compelling talk it looks like we have time for a couple questions and i will take advantage and ask a couple of them myself so first one of the things that that i was really interested in is what type of materials have you and are you investigating for the hardware you mentioned yes so i think that that depends on the application one example that we have used in our lab is that we developed a some number of years ago a material that you can tune its resistivity in many orders of magnitude and this is something that helps you a lot when you are designing analog circuitry because you can essentially make resistors of the same size with custom resistivity moreover these materials can withstand their state to temperatures up to 500 degrees c which means that now you can take these materials and like i said push computing based on uh brain inspired and inside expired approaches into areas where conventional silicone doesn't work great and and sort of a follow-up to that or parallel to that a little bit is what type of applications do you think are best suited for a chip like this so for online learning uh i think you have a wide range of applications like i said one of them of course is self-driving vehicles if you want the vehicle to react to a change that it hasn't been trained for uh without failing catastrophically that's one application another application that is a very unlikely but i think has a could have a lot of potential is in the so-called brain computer interface where you are trying to use smart prosthetics and using the brain to control them and with having an online system means that the burden is not just on the user or the patient but it can also be shared by the the hardware that is trying to adapt to the patient and not the other way around these are just two examples great excellent very very interesting stuff thank you again angel really appreciate the talk today and thank you again to everyone for attending the event today i want to remind you that we will provide written answers to any of the questions that you sent in video via the chat along with the recording of angel's talk uh after this webinar is over if you have any questions at all regarding this or any feedback that you want us to provide please use the email you see at the bottom of your screen there are partners at anl.gov and we look forward to any any feedback that you have thank you again that concludes our our webinar for today and please have a great day you

Channel: Argonne National Laboratory Training

Views: 238

Rating: 5 out of 5

Keywords:

Id: ySotheMYQug

Length: 13min 36sec (816 seconds)

Published: Mon Aug 16 2021