Could AI Technology Be a Source of Advancement?

the agenda with Steve pakin is made possible through generous philanthropic contributions from viewers like you thank you for supporting tvo's journalism coming up on the agenda we even just look at the trajectory of History we have been improving in a lot of key metrics for our Humanity over time but when we look out into the future a lot of the problems that we face today we have solutions for it's just about humans making the right decisions and steering our future in the right direction then science and medicine we use data for all of our tasks to test hypotheses to develop models to make discoveries to make predictions and so Ai and machine learning is helping us use more data data we can use before much more quickly in a much more sophisticated way so the idea is that we can actually take the promise of those that new efficiency and new ability to look at new data and that'll enhance the scientific work that we do faster and perhaps open up new avenues that's ahead on the agenda understand foreign ly flocked to try out a new artificial intelligence app some people pause to ask where does this go now a Canadian futurist who is founder and CEO of the tech education company way has put a lot of thought into what's ahead her name is Sinead bovell she's on the United Nations generation connect Visionaries board and a contributor to Wired Magazine and SHINee bevel joins us now hi hello it's such a pleasure to meet you in person I've been following you on all the social media so it's nice to see you in person you too you as well so you're a futurist what does that mean I think the name sounds a lot more glamorous and mysterious than it is but essentially tracking a lot of data points both qualitative and quantitative and using it to build forecasts and future scenarios so tracking things from emerging Technologies to patents to who's a company hiring and using that to kind of make forecasts about where we could be headed so you mentioned forecasts not predictions absolutely not predictions that's an entirely different wheelhouse so we seem to seem to um living in what media is always saying we're living in unprecedented times and a lot of people I think are feeling maybe less optimistic than they have in the last little while as a futurist what makes you optimistic about the future right so I think if we even just look at the trajectory of History we have been improving in a lot of key metrics for our Humanity over time but when we look out into the future a lot of the problems that we face today we have solutions for uh it's just about humans making the right decisions and steering our future in the right direction but I think a lot of the the key critical problems that we're facing today we can solve them we just have to make the choice to do so and so that's what keeps me quite optimistic and knowing that the optimistic scenarios are very much possible and I think I want to come back to the decisions because Tech seems to be moving at a faster Pace than ever before and sometimes the decisions we make now might not be decisions we make 10 years from now but we'll get back to that in a little bit um you also help Millennials and gen Zed enter the tech space what's the demand been like over the past few years in terms of them stepping into the workforce do they feel more empowered to be in that space I think they feel a lot more comfortable with it um and especially for for Gen Zed technology has been a platform for them to use their voices so I think their approach to it is a lot more inspiring a lot more encouraging a lot I'd say a little bit more optimistic than other Generations do you think we're doing enough to educate people around it because I have uh two small kids and uh over you know the beginning of the pandemic they were online and then before that pediatrician said no screens no Tech and now it seems to be this little like push and pull we relied so much on technology during the pandemic for many different things but now it's like well no we can't um so do you do you think that we know do you think that we're we know enough to make the decisions that we need to make forward or do we just need more education around it I think a lot more education around it and I think for starters a lot of our vision of what technology is is devices but technology is things like software as well so artificial intelligence so Tech education really needs to include those types of concepts for for children as well so they're prepared and it's not just a matter of do we use iPads or do we not should an AI be making this decision about us or if an algorithm put this piece of information in front of us can I think critically to know why it did that those are the types of areas of tech education that I think are important and I don't think we do enough of that you are a Canadian but you work in the states is is it more challenging to enter the tech space here in Canada in the last few years I'd say no I think Canada has really done a great job in opening technology as a lane and really helping some companies uh form and build more of a sector here uh prior to that potentially but Canada is a strong a strong player in the world of technology I don't think we talk about it enough but but we're playing a pretty big role well you did um a tech talk earlier this year or last year in 2022 and you shared this Ted talk about digital avatars right and I'm just going to tell the audience just prepare to have your mind blown we're going to show a clip of it this is shudogram she's a striking South African model likely on the path to a supermodel scroll through her Instagram you can see all of the big campaign she's landed she's been featured in Vogue a few times which is kind of like the Holy Grail and she's also an activist she uses her platform as a rising black supermodel to call for more diversity in fashion and I think that's incredibly admirable there's another fact about shudu she isn't real okay um what yes so shudu is part of an emerging field of digitally generated people so digital Avatar she is empowered by AI she's a digital construction but she books campaigns she's been featured in Vogue quite a few times and avatars and AI as digital humans will play a role in our future fashion modeling spokespeople news anchors will start to see them kind of creep into to more and more Industries you said news anchors and I started getting a little like another caller I mean every every industry is going to have to prepare for the future of work and I think it seems a little bit shocking to see an AI or an avatar in the world of fashion modeling because we previously believe jobs that were more in Creative industries were immune to technology but I don't think any any Lane really is okay but let's unpack this a little bit because um in the world of fashion and you've been a model before in the world of fashion there's not there aren't that many uh women of color in modeling um and they're not they're very few women who are dark skinned so she's booking campaigns which means that she's making money first who's behind shudu and what other I guess complications does this create right so the person who created and who controls shudu is white and male which means the future is heading in a direction where people can create and control identities outside of their own ethnic groups this doesn't inherently have to be a bad thing but it does provide a lot of opportunity for exploitation so in the case of shudu you mentioned Prophet for example Judo represents a real black fashion model but the income her identity generates isn't going to black women it's going to her Creator a white man so we are financially shutting out black women from these opportunities while their identity is still being profited off of and for me this raises quite a few red flags what are the red flags does it raise for you so there's the idea of one being misrepresentation Judo is designed through the lens of a white man's Vision so her skin tone hairstyles all of that it's it's through his version of what he finds desirable so there's a lot of opportunity for stereotyping appropriating and misrep misrepresentation which marginalizes real black women and if you look at this kind of as a as a lens to the Future you can see it playing out as a loophole for companies right so instead of having to invest in diversity or improve company culture a company could just hire or create avatars instead from different ethnicities and kind of manipulate the the relationship those groups may have with that company so those kind of raise some red flags to me and I think the final one is that we have to remember access to the market that creates avatars like shudu and especially more advanced ones that will be powered by AI it's not equal there are certain structural challenges that make it harder for some communities to access resources the time skill the capital to build these types of avatars so we're more likely to see some dominant cultures be the creators and owners and controllers again marginal further marginalizing already marginalized communities AI has been around for a while now why do you think we're more concerned about it now I think we are waking up to we interface with it a lot more I mean social media has been a platform to kind of get these stories out a little bit faster but I think it's it's an intersection of us all just waking up to technology whether it's how our data is being used which which are the what are the main companies that are kind of steering our future we're starting to tune into that and I think it's a really good thing well we've been hearing a lot about chat GPT and Google has its own AI bot uh chat bot called Bard and shares recently um in alphabet Google's parent company sank more than seven percent knocking a hundred billion dollars off the firm's market value when it answered a question incorrectly that was posted in an advert on Twitter um but it still seems like an uncertain time for chat Bots what's your take I think chat Bots are going to be the future of how we largely interface with the internet and with technology I think that they will play quite a dominant role when it comes to Google versus Chachi BT or open AI or Microsoft I think that competition is good for for us and for consumers uh forces companies to innovate and being bring better products to more accountable to Market yes what does concern me with chat Bots is that they aren't actually intelligent right they don't know what they're saying and they sound incredibly scholarly but it could be complete nonsense and when you have companies then competing uh for a first place or to kind of be the first mover it can get sloppy and AI is a serious technology it's not it's not a joke to kind of mess around with and so in this instance where Google Google shop up Bard said something you know incorrect if it had a lot more significance if somebody was using that piece of information to make a critical decision with then that's not so great and it's not so you know inconsequential you mentioned earlier about the displacement of jobs I just wanted to go through some information here for a second and then we're going to talk about that the world economic Forum released a report in 2020 saying that kovitz shifted the way we work the report concluded that the workforce is automating faster than expected and that 85 million jobs will be displaced by 2025 which is now two years away the robot Revolution will create 97 million new jobs but communities most at risk from disruption will need support from businesses and governments analytical thinking creativity and flexibility among the top skills needed data and artificial intelligence content creation and cloud computing will be the top emerging professions and the most competitive businesses will be those that reskill an upscale current employees is there a risk that this move towards automation won't create jobs more equitably absolutely I think the the fact or the point about it creating more jobs is correct I think if we look at historical trend lines technology has net net led to to more more Industries new avenues but how those jobs get created who has access to them and who is given the resources time and skills to Pivot towards them that can absolutely be a problem in terms of equity and if we look in even the last 10 years how income has kind of been divided it isn't trending in a Direction that's favorable or equal and I do think that that's something that we need to be tuning into do you think though that maybe in some ways we're kind of just kind of it's going to happen eventually right so should we not adjust to this uh very real probability yeah I think we the technology isn't going away our future Workforce will be more augmented by it than less so we do need to to lean into it and try to minimize the potential gaps I think if there's anything we can learn from the Industrial Revolution is that you have to take care of your society and your citizens because that's where the real problem Lies when people are left kind of to fend for themselves economically their purpose and otherwise I think we know the result the numbers are in we know which direction we're headed towards we have an opportunity to prepare people as best we can and I think it's really important that we we take that AI goes beyond a chat bot now there are tools out there that can mimic someone's voice and create images that kind of like there's been times when I'm like is this real is this not real there was a recent video of Steph Curry just like you know hitting the the net like getting the basket from all over the court and I thought it was real it was not real um what are the what are the concerns around that around authenticity and transparency and even just maybe like you know the Deep fakes that we're seeing right I think uh deep fakes present quite a critical political geopolitical threat that we don't really have a solution for at this point and the truth about deep fakes is it's not just that we are at risk of believing what's not true but that we stop believing what is true we come become so disoriented in a sea of all this information that we lack a critical discourse in in Direction and I think that that's a real risk and we have already started to see in some you know geopolitical situations deep fakes being used but we were able to kind of detect that there have been companies that have had systems or Microsoft has been a big big player and kind of flagging that but it's very much a cat Mouse scenario we don't have an actual plan and so I do think deep fakes present quite an emerging threat and that should be part of the the tech education curriculum and discussion points how do we spot these how do we know to even look for them in the in the first place what about voicemail making that as well I would kind of put that so there's deep fake audio deep fake visual I would kind of lump that all in together and the technology is just further advancing so you had mentioned GPT there are similar Technologies where you can take a three second clip of somebody's voice and use it to generate an entire podcast that they never were featured on and so on the one hand it's going to be great for creators and all of these new tools and resources but on another end it's the disinformation and misinformation can be produced at a fraction of the cost and at a fraction of the speed and that I think is quite an emerging threat that we need to tune into um that's a very nerve-wracking did you uh recently when you were talking about uh we're talking about the mimicking of the voices but when we talk about art you know art comes from a place of experience and emotion and someone's perspective but recently there was this uh I guess not controversy but more conversation about this one app that could create images of all of us that were just stunning where what are your thoughts about that like the AI in this space is kind of in art and creating works of poetry and literature so I think there's definitely been a lot more hesitance or kind of pushback on AI stepping into these creative Realms and for many reasons I think up until this point we thought creativity was something that's uniquely human and to see it be synthesized by an AI system it's very alarming because it changes how we relate to ourselves as humans right poetry art music those are things that we think only some humans are even naturally endowed with and to see it get passed to a machine seems very jarring but I think every single industry is going to be impacted by technology and the world of the Arts is is no different and I'm looking forward to what could be Unleashed uh in an optimistic way when we all have access to creative tools I do think if you are more naturally endowed with artistic skill you could use an AI system much better than somebody who can't paint or or make music so I don't think it's that you know artists go away they can you know adopt these tools for themselves but I think that there will be some magic to to happen when we all have access to these systems to help bring different types of content or creative expression to life is that the optimism that you were talking about that one day possibly we would all have access to these tools yeah I think that would be the goal that we could all kind of have access to them I think of course there are risks that come again with with everybody of access having access to these types of tools but I do think it could be quite a great turning point for society if we use them and adopt them correctly and we're in formed and how to utilize them well there's been more effort on AI regulation in recent years the AI act in Europe those potentially concerned about chat Bots being used in class assignments Etc what are your thoughts and Regulation and reaction to emerging Tech so there is the when it comes to chat gbt in the recent chatbots some schools have leaned to just ban it outright and I think we're the schools might ban it but the kids will find a way to use it the kids are going to use it anyways and I think we're moving in the wrong direction because the purpose of education is to prepare students for the economy of tomorrow and that economy is going to be largely underscored by technology such as chat gbt and other AI systems so we really do need to be equipping kids with not even just the skills to utilize these tools so they can actually be productive people in the economy but to utilize them safely right so when we have conversations around misinformation and disinformation and the risks these systems present if we're just Banning them we're Banning an opportunity for the Next Generation to adopt these tools wisely and steer their future in a direction that they want to use it or for it to go sorry and I think most of our current education systems are largely transitioning or encouraging kids for jobs of the past not transitioning into the jobs of the future and that means we've got to adopt these Technologies we've got to lean into them and we have to also equip students with the skills to build them as well to be in instead of being um I guess content instead of just being the users actually be the creators be the creators and the critical thinkers when social media became a part of our Lives I think a lot of us didn't realize that putting so much of our personal information might come back later to bite us uh parents sharing photos of their children I've done this and then I did it without consent when they were younger and if you are younger you might share something that decades later might impact your employment and sometimes scammers only need your email address to destroy your life should all jurisdictions be following the eu's lead with their right to be forgotten Privacy Law absolutely I think data for many reasons uh could be a national security crisis uh having a bunch of having citizen data just open manipulated accessible to not only just different companies but different countries uh I think that that's a really big red flag and then of course for for personal reasons you should have the right to be forgotten or for a company not to be able to make a statistical prediction as to what your next moves are going to be because they've been hoarding a series a bunch of information on you throughout your life so I think that you is definitely moving in the right direction and I think companies would or countries so it would be wise to follow suit do you think that maybe this is in part why some people are skeptical of AI technology and maybe this future of automation because it just kind of feels as if you're the user but not you're not in control of what's happening around you I think the the fear and kind of pessimism towards AI in the future there's a few reasons some of it's pop culture related tuning into shows like Black Mirror The Matrix um but then again yes of course a lot of our interactions with technology uh we feel like it's happening to us uh not with us and for many people it seems like there's five or six companies uh and maybe seven or eight people kind of steering our future and that feels quite alarming but I think if you equip people with the right information and tools to participate in creating the technologies that they're going to be using I think that can change kind of the discourse in in the direction and at the beginning of the pandemic work for a lot of people changed a lot of people if you were able to you were able to work from home kids were learning online as I mentioned what would you say are the upsides to the future of work right so I think the pandemic showed us who were we incorrectly looting from the workforce because we falsely assume that everybody had to show up nine to five Monday to Friday that opened up a whole new door for for new parents or people with different Mobility needs so that was a kind of key Shining Light I think the future of work is going to be a lot more flexible so we've already been moving away from the days where you work for a single company for 10 years and that trend is going to continue the pandemic and Technology showed us that flexibility can work and it kind of prepared us for what more remote and more transient kind of workplaces look like and I think that that is something that's that's optimistic and helpful well from your view I think for employees it works to have a little bit of Leverage now but from the perspective of the employers how are they taking that future of work are they adjusting well are they going to fight it maybe I I guess it kind of depends on the company and I think we are adjusting well as employees but I think we don't even fully realize what's coming so when I say more remote work or um what I'm really referring to is in a world where smart machines learn new tricks over time it becomes much less likely that a company is going to hire for a full-time role if that role is going to be radically changed in the next year or two years so we're going to see a rise of of the gig economy across all jobs so we see a lot of you know whether you're a delivery you know driver or whatnot we'll see it across Financial analysts lawyers Physicians teachers where we all work in kind of different roles for a few different companies at any one time and in terms of Are We embracing that our company is embracing that I don't think we're we're ready yet but I think things like remote work have been helpful in kind of laying the foundation for how those systems and infrastructures could operate a couple of years ago you wrote how we need to stop asking kids what they want to be when they grow up what did you mean by that and I know that that question is usually asked with the best intention but the reality is most of the jobs that a child would see today or or answer that question with probably won't Exist by the time they are in a working age or they'll be radically transformed so instead of kind of setting people up and it's not for for failure but um for the idea that your identity is attached to the job that you do when we know the future of work is going to be very different we should be encouraging kids to think more broadly about the problems that they want to solve because because those are are more likely to exist than a specific occupation and especially as technology continues to more rapidly disrupt the workforce we have to move away from this idea that you are your job because that's going to change so if we can start with kids and encouraging them you know what are the skills that you want to learn and the problems that you want to solve and one of the most important skills for the future being critical thinking and Imagination that's something that children are already naturally endowed with lean into that the Curiosity when you said that I kind of felt like oh I think that that much that's hard for people to hear it is but you know 15 years ago the role of a social media manager analyst didn't exist now if a company doesn't have one you're probably not going to make it uh so the trend lines aren't very different there are significant and you know whether you're a data analyst a data scientist all of these jobs have really come to the Forefront in in the last decade and so that's still going to continue going forward but I think the future always seems a lot more shocking from the present and especially when we analyze it through the Frameworks of the present how can we Empower then people to embrace technology and to leverage it I think leaning into it you know the the best thing we can do about the future is prepare for it it's not going to go away um to recognize how much you already use it I don't know about you I can't get down the street without consulting maybe Google Maps uh so we use AI all of the time social media if you watch a streaming platform so to know that these tools aren't as overwhelming as we might make them out to be they're very actually easy to to use but it's about leaning into it doing your best to try to understand it and keep up with the discourse with it and then of course at a more societal and federal level level equipping people with the resources they're going to need to thrive in such a dynamic future Sinead it's been amazing having you here I I could talk to you for another hour thank you so much continued success to you thank you all of us thank you the arrival suddenly of artificial intelligence in people's everyday lives has unsettled even those normally very bullish on new technologies but what if the raw power of this technology could mean a finding potential cures and treatments in weeks rather than years or decades that's already happening so let's find out more with in Palo Alto California Daphne caller co-founder and CEO of the biotechnic company in citro and an adjunct professor of computer science and pathology at Stanford University in Atlanta Georgia Annette matabushi professor of biomedical engineering at the Georgia Institute of Technology and Emory University and here in our studio Laura Rosella professor and Canada Research chair in population Health analytics at the University of Toronto an education lead at temerity Center for artificial intelligence research and education in medicine welcome Laura thank you studio and to those on the line so last month more than 1800 people signed a letter including Elon Musk and apple co-founder Steve Wozniak calling for a six-month pause on the development of AI Laura I'm going to start with you many are clearly concerned about the downside of AI but how would you characterize the promise of AI in advancing science especially in medicine and in healthcare yeah well in science and medicine we use data for all of our tasks to test hypotheses to develop models to make discoveries to make predictions and so Ai and machine learning is helping us use more data data we can use before much more quickly in a much more sophisticated way so the idea is that we can actually take the promise of those that new efficiency and new ability to look at new data and that'll enhance the scientific work that we do faster and perhaps open up new avenues all right same question to anant how would you characterize sort of the promise of AI and advancing the medical field yeah thank you Jan I think the opportunity is tremendous for AI in health and Medicine particularly when you think about some of the big challenges and problems that we face today problems around Health disparities problems around Global Health particularly in Low Middle income countries areas of the world where there is not a pathologist or a radiologist in the entire country and that's where I see the big opportunity for artificial intelligence to really make a difference in those areas of the world that really are underserved and don't have access to medicine and health care and AI really being that great leveler Daphne is AI ushering in a new era of science I think it is because as Laura said we are in a world where we have the opportunity to collect a tremendous amount of data about human biology and human biology is incredibly complicated far too complicated with the human mind to fully understand and um and many of the diseases that remain with unmet need today are probably not ones that these and we're probably not characterizing them correctly Alzheimer's is not a single disease and if we're able to collect enough data about human biology and use the power of machine learning and artificial intelligence to disentangle what we see there maybe we can finally characterize the right subsets of patient population and find interventions that have a meaningful effect size and that would be transformative with that I want to look at some recent advancements in medicine using AI science magazines 2021 breakthrough of the year was powered by Ai and its ability to predict protein structures in the body last month researchers used artificial intelligence to detect Alzheimer's risk with over 90 accuracy and earlier this year researchers used AI to discover a potential new cancer drug in less than 30 days in what usually takes years or even decades Laura I'm going to come to you machine learning and AI have been around for a while what is happening now that is enabling these breakthroughs you sort of mentioned it a little bit off the top data yeah we have the data that we just did not have before so the fundamental you know mathematical basis of of AI machine learning has been around for a while but we did not have the amount of data that we have now in the granularity of that data and then we didn't have the computational ability to process that data so theoretically we could do this but now we actually practically do it we have the horsepower behind it to actually make these discoveries happen and that's been the biggest change this is something that we harp on here in this country is that in Canada we don't have a lot of data where are we getting our data from is this a publicly sourced data is where are we scouring for it yeah the data come from lots of different sources so from the biological point of view data can come from the cells and what we measure at the cellular level and the biological level for data I work with on humans we get it from interactions with the Healthcare System we're doing more and more surveys and detailed measurements it can come from non-traditional sources like social media or our devices and so it's actually coming from all these different places I work with environmental data so it's coming from the sensors that measure things in the environment images in the environment so it's not coming from one place it's coming from many places and it's very variable that's the hardest person to say is there some biases to this with Sciences very methodical in terms of how we get our data is is there bias in some of the data that we have yes there's a lot of advice and this is the part that makes people probably the most uncomfortable the data the data generating processes are non-random when we work with this data I as an epidemiologist obsessed with bias a lot so we think carefully about what the biases are how we can mitigate it against them and of course having more data independent replication making sure we're not just working in one Center but multiple centers these are some of the ways we overcome some of the biases all right and ants I want to come back to you I want to go back to uh the the recent advancement and sort of the excitement what areas of research are you seeing the most exciting results using AI so I think that there's been a lot of progress around AI for Diagnostics well I think there is something like 300 technologies that are now been approved by the FDA in the United States for primary diagnosis of multiple different indications there's a lot happening in the realm of Ophthalmology there's work that's going on in the Cardiology space but looking forward I think Beyond disease diagnosis I think the opportunity is tremendous when it comes to predicting outcome when it comes to predicting therapeutic response and I think that that's where I see the next Frontier where we in the United States for instance 40 of the adult population is going to be diagnosed with some form of cancer in their lifetime and the big question is how do you manage these patients we know that you can't be hitting all these patients up with aggressive treatments I think a big question for us today is better management for patients with a disease and I believe that the next Frontier will be the use of AI to help in tailor the appropriate treatment strategies and management options for patients going forward Daphne my natural follow-up is how is AI changing the discovery of new drugs uh to treat said range of diseases so I think first and foremost is identification of the right patient population because I think we have some really compelling treatments today but when you apply them at the population level you have some subset that responds and a very large number that do not and for those that do not respond you've basically introduced a tremendous amount of toxicity you've prevented them from uh benefiting from a treatment that might actually help them so I think disentangling the complexity of disease to understand what are the right population subsets that you can create a therapeutic intervention with a meaningful effect size is a critical part of it and then the next part is identifying intervenable causal nodes that if you actually modulate those nodes those proteins or other metabolites in the body it will actually make a difference to those to that coherent group of patients and we're seeing a tremendous amount of development on the drug Discovery side in first deconvoluting the patient population interrogating causality and then um and then figuring out how to construct chemical matter of whatever whether it's a small molecule or a large molecule or a gene therapy that would actually make a meaningful difference in intervening at that node so I think there is just progress all throughout that that process that's happening and is being driven by AI Daphne what have you come up with in terms of that research there so in our work um first of all I want to return to Laura's point we um take very great care in how we collect and generate our data so in addition to collecting data from Human populations we also have a cell Factory that generates um stem cells and uh from Human derived populations patients um healthy controls so that we can actually measure disease at the cellular level and um and so we have done some really exciting work in uh in various seasons of Neuroscience as well as in oncology and uncovered what we believe are a compelling new class of targets for for genetic epilepsies which is the first therapeutic area that we moved into as well as in the domain of cancer identified both new targets as well as importantly new patient populations for drugs that already exist are quite safe are effective but are not deployed in the right patient populations and so those are the fastest path to clinical impact for for getting into patients because the path of getting from a Target to an approved drug is quite a long one working with an existing drug is sometimes a much faster path all right Laura I'm gonna come to you uh we're gonna get granular in our research here you used a machine learning model to analyze Health stat of 2.1 million people living in Ontario what were you able to discover we would we were able to accurately predict who among that population would develop diabetes in the next five years type 2 with diabetes and the idea behind it is if we understand who can develop type 2 diabetes who is likely to we can prevent it type 2 diabetes can be prevented we have uh well-known interventions that can prevent the onset or delay the onset of type 2 diabetes so if we know who's most at risk we can Target those interventions more appropriately how accurate can we get can we get to the individual with this this is a good question I mean no model really gets to the individual level per se because what machine learning does and what all of these models do is say someone that looks like this with these characteristics on average this is their likelihood of developing an outcome so we can be very accurate for an individual certainly but every individual will have slight nuances for sure but much more accurate than me just trying to guess oh yeah you have one two or three risk factors you're probably going to develop diabetes we can then use hundreds of variables much more nuanced information we're going to get much closer to our ability to predict the fact that you will or will not develop diabetes all right my next question it's a dangerous one because I'm we could get in the weeds here for sure but how were you able to do that exactly is there an example that we can sort of understand of how you were able to sort of get hone in on on these on people who were at higher risk yeah I mean so simplistically the way these models work is they go through all these variables and they lump together characteristics that are most likely appearing in individuals that do develop diabetes and do not and they essentially sort through and trying to identify common patterns between those individuals it's a lot more complicated than that but we take in information on the history past interactions with the health system other conditions they might have other risk factors their age and so putting all these things together you can come up with a score to determine you know based on all these factors put together this person is likely to develop an outcome and in this case we know what we're trying to predict which is type 2 diabetes but sometimes these methods can be used where we don't know we're just trying to group some groups together that might be more similar than others and in that case it's much more of a discovery angle saying we know there's some commonality here we don't know yet what to do but we can try things and I think that's some of the work that you heard about very interesting all right and ants I'm going to talk to you about another disease cancer how have you been able to use AI to improve outcomes for people with cancer so our group has been really interested in how we could use AI with routingly acquired data so talking about radiologic scans CT scans MRI scans but also pathology images to be able to really figure out which of these patients has the more aggressive variant of the disease and therefore is going to benefit from more aggressive treatments like radiation or chemotherapy vis-a-vis those patients where perhaps the disease is not quite as aggressive and some of these patients might benefit from perhaps no chemotherapy or a lower dosage of radiation therapy or in some cases like in the case of prostate cancer there are several men who might benefit from no intervention at all because we have a much less aggressive variant of the disease and so there's a big opportunity that we're finding with AI with machine learning to be able to tease out patterns from these radiologics samples images from the pathology images to be able to really help wrist stratify those patients who need the more aggressive treatment regimens versus those patients who might benefit from a less aggressive variation of treatment strategies because they have a more indolent or less aggressive cancer Daphne in your experience with drug Discovery to what extent is AI helping us discover things that human scientists could never discover so I would like to uh um begin where an aunt left off which is with the incredible amount of information that exists in uh in images such as histopathology and radiology and in many of those cases when we apply machine learning to those images we uncover patient populations that humans would never have been able to identify subtle patterns within the histopathology images which is incredibly information rich source and also collected abundantly because pretty much every solid cancer patient has that has a biopsy taken as part of the standard of care and so we've been able to take those large populations within say breast cancer within colorectal cancer and identify patient subgroups where particular Gene is driving seems to be driving the cancer in ways that are not necessarily the case for other patients that have what is labeled as the same cancer as part of their medical record and that driver Gene is now a novel cancer Target but importantly neither the patient population nor the fact that that Gene is driving the cancer would have been identified by a human pathologist because the patterns there are too subtle for a person to have necessarily picked up on now what's interesting is that sometimes once we've discovered that we've developed methods for visualization for explainability so that we can show the pathologist what it is that the cancer is picked up on and that's important because it also allows us to gain confidence that what's been discovered is not artifactual it's it's actually real it's not some kind of a biases in the data as Laura was alluding to earlier but it also helps teach the Pathologists something new about um about tumors and and how the cancer is different in in different populations so it's um so it teaches us new biology at the same time that it uncovers potential new and impactful intervention nodes all right with that let's talk about some of the challenges here and to get into that I want to read a quote from a Scientific American AI is currently hampered by a lack of transparency this lack of transparency has been nicknamed The Black Box problem because no one can see inside the network to explain its thought process not only does this opacity under my trust in the results it also limits how much neural networks can contribute to humans scientific understanding of the world first off Laura why can't we see into this black box well the black box is an issue but I actually think we need to get much more specific about what we need to do and and Daphne alludes to this as well so to me there's at least three parts of the Black Box there's interpretability we need to know what's in these models explainability how are the different components contributing to how the model makes decisions and then there's transparency which is a big one I would say that's the one that's probably most problematic nowadays because it's done so variable um and that is the way the model is constructed needs to be very clear clearly documented standardized it needs to be reproducible someone needs to be able to independently verify it and so right now none of those things are happening consistently and all of that's contributing to the black box and so once we do have those things in place I think we'll feel a lot more comfortable it's not as simple as oh I just need to see look under the hood and see what's in the model and all all is okay all those aspects actually need to be working for us to feel more comfortable in terms of the black box and and sort of ai's contribution when a scientist goes in are we getting a step-by-step in terms of how it got to that result not consistently and some of it we can't fully explain so some of it is something that oh this is new this is a pattern I wasn't expecting we haven't seen this before which is not necessarily a problem but it means it has to be verified someone else in another institution might want to verify that we might want to check it we might want to do further experimentation so it's just a step one but it's it's pretty complicated it's not as simple as just saying oh these are the variables in the model and stopping there all right we're going to pick up on that a little later but I want to bring Daphne into this how big of a challenge is the Black Box problem I think it both is and is not so there's technique that people have developed including ourselves to help people see into the black box and and sort of interpret what it is that the computer is picking up on and it's especially easy in the context of images because you can visualize um different variables and what it is in the image space that um that give rise to different conclusions and so that gives I think people a much greater sense of confidence about what is going on I think the other aspect which is important to emphasize in in drug Discovery specifically is that where the hypothesis came from matters up to a point But ultimately you're going to be subject to the same rigorous test that every other drug is subject to which is a randomized clinical trial and so in the same way that we don't necessarily ask people where a particular drug hypothesis came from when we go into the clinical trial phase because we know that there is a sort of fundamental ground truth verification process which is does this work on a randomized case control popular and I think that's a really important component that um that gives us confidence in any drugs that we put into a human an answer in general if we can't understand or see how AI has arrived to some potential promising at some potential promising results uh you know with other words of confidence and transparency how can we trust those results yeah so so thanks for that Jay and just to keep things interesting I'm going to be a little contrarian and and maybe disagree a little bit we're Daphne there so I can give a very specific example of where the black box really came back to hurt us about six years ago we were working on a project looking at the use of AI to predict heart failure from endomyocardial biopsies and we train the network uh we were able to demonstrate from a single institution that this network was able to predict the risk of heart failure based of what the network had learned or what the the black box had learned from these biopsy images and we were stunned by the results we actually showed in in a paper that the Black Box outperformed cardiac Pathologists by 25 we were ecstatic but the PostScript to that story is that a few months later we got another tranche of images from that same institution and when we ran the network again the performance fell from the 97 that we'd achieved on the first pass to a 75 result which was more in line with what the pathologists had been getting and we found out only after much interrogation that what had happened between the first launch of images and the second launch of images was that a remote software upgrade had been applied to the scanner that had been used to digitize the slides and that very subtle change in the appearance of the images had perturbed the network enough that it had drastically changed its performance from the 97 to down to 75 and so because of that experience our group has really focused on very intentionally interpretable approaches and you know while we talk about the promise of AI to discover things that we don't know one of the things to consider in all of this is that there is a huge body of knowledge that has been amassed for many diseases over the course of of decades and potentially hundreds of years let's take histopathology as one example that you know Daphne had referenced as well you know Pathologists have been looking at slides under a microscope for over 100 years and there's a deep well of knowledge that we have about the kind of Hallmarks and patterns that are associated with more aggressive and less aggressive disease the challenge is that Pathologists don't do a very good job in terms of reproducibly identifying these Hallmarks and that's where I think the AI can be so powerful because if the AI could help identify in a reproducible fashion these patterns these Hallmarks then we are able to have very intentionally interpretable AI from the get-go and can demonstrate its relevance for diagnosis prognosis and treatment response prediction yeah Daphne if you want to respond yes so I can I can if we just focus on reproducibly and reliably doing what a pathologist can already do we are missing an incredible opportunity to discover new Concepts and I will refer back to a paper that I wrote back in 2011 which was one of the first machine learning analyzes of histopathology data which basically discovered the importance of the tumor microenvironment to cancer prognosis at a time when we did not realize that as a community so it was one of the first earliest harbingers of the importance of the tumor microenvironment specifically because uh we did not restrict the model to trying to do something that replicated what Pathologists already doing so that being said I completely agree that this needs to be done with care there needs to be replication ideally across distinct hospitals not just doing with a verification within a single hospital system in a single scanner device but really replication across ideally multiple cohorts um there needs to be an explainability component trying to help us understand what the machine is picking up on and as I said in the context of drug Discovery the ultimate proof is a randomized clinical trial which is really hard to um to fake so I think I agree that we need to do this with very great care and that it's easy especially with modern day machine learning to fall into a trap of uh of getting to ridiculously high performances that are not driven by anything that is biologically meaningful completely agree with that but I don't want to throw out the baby with the bath water in preventing the machine from identifying new and important insights that a person did not previously discover all right moving on Laura we know AI can detect patterns and make predictions when it's got massive amounts of data we know it's limited in sort of explaining why something is happening and what is causing it does this mean that AI will never replace human scientists there's a fear out there with every industry right yeah AI is coming for us but in this with science this is a this is a different realm but the amount of progress that's been made is quite tremendous yeah I don't think of the term replace when I think of AI and lots of people go there right away I always think of augmenting and so I don't I think it will augment the work of scientists for sure and causality and understanding why things happen is a great example there are important parts of that process it's a multi-step process and some pretty key steps that need to be taken to really understand why and machine learning can play a role in some of those steps and actually can help us get to that more quickly it can open us up our eyes to new possibilities that we weren't seeing before but we still have to go through all the steps so I don't see it replacing I mean I'm a little hesitant to make any predictions in the space given the progress we've seen in the in the past year even but I see it much more as augmenting the work that scientists do and you know still making sure that we keep the rigor and the domain knowledge that we've amassed over the years in that process I want to get your take on that as well yeah so I think that the the the these these Prophecies of Doom are not new right so AI is going to take away the clinician's job AI is going to replace the pathologist AI is going to replace the radiologist as a history buff I've gone back and looked at some of the stories that came out you know 30 years ago in the context of digital mammography when digital mammography came out in the late 80s or early 90s and some initial image analysis machine learning companies were stood up to analyze digital mammography you could see these Prophecies of Doom you know you're not going to need Radiologists anymore you fast forward three decades on we still have a paucity of Radiologists we have a shortage of Pathologists so I agree a hundred percent that we really need to be thinking about augmentation we need to be thinking about assisting and I think that one of the one one of the things that's been said more recently about AI is that yes it is likely that in the next five to ten years the Radiologists of the pathologist who is using AI might obviate the radiologist or pathologist who does not Daphne could AI develop to the point where it can create its own genuine scientific understanding I I might be going a stretch here but think like an AI Einstein I definitely be wary of predictions saying AI will never do X Y and Z because in the past whenever we have made such predictions um they have turned out to be in many cases fall so I don't know if never but honestly I take I tend to agree with uh with Laura that scientific Endeavor is one of the places where it's the ultimate creativity of the human mind and it is certainly the case that a human can be hugely um supercharged if you will by having access to the kind of interrogation tools that allow a human to extract patterns um and insights from very large amounts of data which the human mind will never be able to uh to process but I think that the that the partnership between the human and the computers where we're going to see the greatest insights now will a computer come up with insights on its own probably but I still think that the greatest insights are going to come from that partnership all right Laura it's been said that when Einstein made his big breakthrough with relatively only a handful of people in the world actually understood it could AI or an AI Einstein produce a scientific breakthrough that no human could understand so I think that an AI could generate a starting step of a discovery that we will eventually understand um so I think that signals will come up some of them will be false some of them will be because of a bias or spurious finding and we have to do a process to get there and some of them truly will be New Paths of understanding that we did not yet have so I think it could start us on that path and I I'm confident though that eventually with the processes that we use in science we will eventually understand it might take us time though so it might be signaling us something that we don't yet understand but I'm confident that we will be able to get there in time all right we are going to leave it there Daphne and Aunt Laura thank you so much for joining us on the program very very riveting stuff thank you so much thank you thank you thank you [Music] I'm Jane Jacqueline thanks for watching TVO for joining us online at tibio.org and we'll see you again tomorrow the agenda with Steve Pagan is made possible through generous philanthropic contributions from viewers like you thank you for supporting tvo's journalism [Music] foreign

2023-09-18 12:57

Other news