Transcript: Can AI Combat COVID-19? | Mar 29, 2021

Steve sits in the studio. He's slim, clean-shaven, in his fifties, with short curly brown hair. He's wearing a gray suit, white shirt, and spotted gray tie.

A caption on screen reads "Can A.I. combat COVID-19? @spaikin, @theagenda."

Steve says THERE ARE NOT MANY SILVER LININGS IN THIS PANDEMIC. BUT ONE OF THEM IS THAT, IN LESS THAN A YEAR, NOT ONLY DID SCIENCE MANAGE TO DECODE THE VIRUS, BUT MULTIPLE VACCINES WERE ALSO DESIGNED, TESTED AND ARE NOW OUT THERE MAKING A DIFFERENCE. THAT'S JUST ONE EXAMPLE WHERE THE USE OF ARTIFICIAL INTELLIGENCE HELPED SPEED UP THE FIGHT AGAINST COVID-19. WITH US NOW FOR MORE ON HOW AI IS HELPING: FROM THE PROVINCIAL CAPITAL, IN RIVERDALE, WE WELCOME: NAHEED KURJI, CO-FOUNDER OF THE BIOTECH FIRM CYCLICA, THAT USES AI TO ACCELERATE DRUG DEVELOPMENT...

Naheed is in his early thirties, clean-shaven, with short black hair. He's wearing a gray sweater and a black shirt.

Steve continues AND IN THE ANNEX: MARZYEH GHASSEMI, ASSISTANT PROFESSOR OF COMPUTER SCIENCE AND MEDICINE AT THE UNIVERSITY OF TORONTO, AND THE CIFAR AI CHAIR AT U OF T's VECTOR INSTITUTE CIFAR BEING THE FORMER CANADIAN INSTITUTE FOR ADVANCED RESEARCH...

Marzyeh is in her thirties. She's wearing a black hijab.

Steve continues WE ARE DELIGHTED TO HAVE YOU TWO WITH US ON THE PROGRAM TONIGHT, WITH AN INTERVIEW ABOUT WHAT THE HECK YOU TWO ARE UP TO, BECAUSE IT'S VERY FAR-REACHING, FUTURISTIC STUFF THAT'S HAPPENING RIGHT NOW IN THE PRESENT. MARZYEH, GET US STARTED HERE. WHAT ARE YOU ACTUALLY WORKING ON RIGHT NOW?

The caption changes to "Marzyeh Ghassemi. University of Toronto."
Then, it changes again to "Current developments."

Marzyeh says I THINK A LOT OF THE THINGS I'M WORKING ON ARE HELPING PEOPLE REPORT THEIR SYNDROMES ONLINE. THAT IS ONE WAY WE CAN SUPPLEMENT WITH TESTING. IT'S HARD SOMETIMES TO GET A TEST IN A REASONABLE TIME FRAME, AND SO IF PEOPLE CAN SELF-REPORT HOW THEY'RE FEELING, THAT CAN GIVE US A BETTER IDEA OF HOW THE PANDEMIC MIGHT BE GOING, ESPECIALLY WITH THE NEW VARIANTS. AND I'VE ALSO BEEN WORKING ON COLLECTING A LARGE NUMBER OF CHEST X-RAYS FROM PEOPLE IN DIFFERENT COUNTRIES THAT HAVE COVID-19 AND COMPARING THAT TO PEOPLE WHO HAVE PNEUMONIA. AND THAT WAY WE MIGHT BE ABLE TO DO BETTER EARLY ON WITH PREDICTIONS, ESPECIALLY WITH A NEW VARIANT, TRYING TO SEE IF THERE IS A BIG DIFFERENCE.

(baby babbling)

Steve says DO YOU HAVE A LITTLE VISITOR WHO WANTS A CAMEO AT THE SAME TIME HERE?

Marzyeh says I DO. I DO. YOU KNOW, WE'RE ALL VIRTUAL SCHOOL, EH? I HAVE MY YOUNGEST HERE.

Steve says YOU CAN FEEL FREE TO INTRODUCE AT ANY TIME, IF YOU WANT. WE UNDERSTAND.

The caption changes to "Marzyeh Ghassemi. Vector Institute."

Marzyeh says I PROBABLY WILL AT SOME POINT. I THINK THAT THE LARGEST THING THAT I HAVE BEEN TRYING TO FOCUS ON ARE SOME AREAS WHERE MACHINE LEARNING CAN ACTUALLY HELP. SO, FOR EXAMPLE, IN EARLY RECOGNITION OF SOMETHING LIKE COVID IN A CHEST X-RAY, IF WE'RE IN A RESOURCE CONSTRAINED SETTING, OR IN SELF-REPORTING OF SYMPTOMS WHERE WE KNOW WE CAN'T GET COMPLETE TESTING COVERAGE.

Steve says UNDERSTOOD. OKAY. WE'LL FOLLOW UP ON THOSE TWO THINGS AS WE GO ALONG HERE. NAHEED, WHAT IS CYCLICA AND WHAT IS ITS MISSION?

The caption changes to "Naheed Kurji. Cyclica."

Naheed says THANK YOU FOR HAVING ME. WE ARE A BIOTECH COMPANY THAT DEPLOYS NOVEL METHODS IN COMPUTATIONAL TECHNIQUES LIKE AI TO SPEED UP THE WAY IN WHICH MEDICINES ARE DISCOVERED, AND IT'S A VERY EXCITING SPACE AND THE WORK THAT WE'VE BEEN DOING IN COVID VIS-A-VIS THE APPLICATION OF AI HAS BEEN PRETTY INTENSE OVER THE PAST 12 TO 14 MONTHS.

Steve says JUST FOLLOW UP ON THAT, IF YOU WOULD? HOW DO YOU ACTUALLY USE ARTIFICIAL INTELLIGENCE TO HELP US MAKE PROGRESS AGAINST COVID-19?

Naheed says YEAH. SO LET ME QUICKLY TALK ABOUT A PATIENT JOURNEY. WE'RE ALL PATIENTS. EVERYBODY LISTENING TO THIS IS A PATIENT. AND WHEN WE GO TO THE DOCTOR AND THE DOCTOR, YOU KNOW, DIAGNOSES US WITH A CERTAIN TYPE OF DISEASE, WHETHER IT BE COVID OR SOMETHING ELSE, YOU KNOW, THEY'RE GENERALLY TRYING TO UNDERSTAND THE DRIVERS OF THAT DISEASE. AND THEN RESEARCHERS FOR THAT SPECIFIC DISEASE ARE FOCUSED ON DESIGNING MEDICINES TO COMBAT IT. AND GENERALLY THE MIDDLE POINT OF DESIGNING A MEDICINE FOR A DISEASE, THEY'RE FOCUSING ON DESIGNING A MEDICINE FOR A PROTEIN. A PROTEIN IS THE BIOLOGICAL DRIVER OF THAT DISEASE. AND WE USE AI TO SPEED UP THE WAY IN WHICH MEDICINES, CHEMISTRY, INTERACT WITH THE BIOLOGY, THE PROTEINS, FOR A GIVEN DISEASE. NOW, WHAT WE'VE DONE IN THE COVID ENVIRONMENT, STARTING OFF AT THE END OF JANUARY OF 2020, WE KICKED OFF A COVID TASK FORCE AT CYCLICA TO UNDERSTAND OUR CONTRIBUTION OF USING OUR AI TECHNIQUES FOR REPURPOSING DRUGS THAT EXIST FOR COVID BIOLOGY. THAT TASK FORCE THEN PUBLISHED SOME PAPERS IN EARLY FEBRUARY ON A DEEP LEARNING AI BASE RESOURCE TO IDENTIFY REPURPOSED DRUG CANDIDATES FOR COVID-19. WE THEN OPENED UP A COVID STIMULUS PLAN, WHERE WE ENGAGED WITH THE ACADEMIC COMMUNITY GLOBALLY, AND WE KICKED OFF OVER 20 COLLABORATIONS TO USE OUR TECHNIQUES FOR THAT SPECIFIC REASON, AND ONE OF THEM WAS A COLLABORATION WITH THE UNIVERSITY OF TORONTO THAT WAS CIFAR FUNDED, AND IN THIS PROJECT THAT WAS LED BY CIFAR GLOBAL SCHOLAR FROM THE UNIVERSITY OF TORONTO SICK KIDS, WE AND A GROUP OF RESEARCHERS USED OUR AI PLATFORM TO IDENTIFY EXISTING SAFE DRUGS THAT COULD BE REPURPOSED TO TREAT COVID-19 PATIENTS. AND WHAT WE FOUND WAS, WE NARROWED IT DOWN TO A LIST OF 15 TOP DRUG CANDIDATES AND WE'VE IDENTIFIED AT LEAST ONE APPROVED DRUG TO DATE, WITH MORE PLAUSIBLE DRUGS THAT CAN BE REPURPOSED FOR TREATING COVID-19, WHICH ACTIVE WORK IS UNDERWAY. SO IT'S BEEN REALLY EXCITING AND, AGAIN, FUNDED AND SUPPORTED BY CIFAR WHICH HAVE DONE SOME INCREDIBLE WORK SUPPORTING INNOVATION OVER THE PAST 12 TO 18 MONTHS.

The caption changes to "Watch us anytime: tvo.org, Twitter: @theagenda, Facebook Live, YouTube."

Steve says ONE QUICK FOLLOW-UP. WHAT DO YOU HOPE THOSE DRUGS ULTIMATELY WILL BE ABLE TO DO?

Naheed says THERE ARE TWO THINGS. ONE THAT IS IMMEDIATELY RIGHT IN FRONT OF US WITH COVID. AND WHILE THERE ARE VACCINES BEING ROLLED OUT, THERE ARE STILL THERAPIES THAT ARE REQUIRED FOR THOSE THAT GET SICK. SO NUMBER ONE IS, WHAT IS THE IMPACT THAT WE CAN HAVE IMMEDIATELY TO ATTENUATE THE ILLNESSES THAT ARE BEING EXPERIENCED. AND NUMBER TWO IS, WHAT CAN WE LEARN FROM THIS ENVIRONMENT TO LESSEN THE SEVERITY OF A FUTURE INFECTIOUS DISEASE THAT IS NOT ANTICIPATED, LIKE COVID WAS. SO HOW CAN WE COME TOGETHER, AND WHAT ARE THE METHODOLOGIES THAT WE CAN BRING TOGETHER IN COLLABORATION WITH OTHER INSTITUTIONS TO MAKE THE INNOVATION PIPELINE MUCH FASTER.

Steve says GOTCHA. OKAY. EXPECTATIONS OF WHAT ARTIFICIAL INTELLIGENCE IS CAPABLE OF THAT WE DON'T NECESSARILY UNDERSTAND THAT THERE ARE ALSO SOME OBVIOUS LIMITATIONS ABOUT WHAT IT CAN AND CAN'T DO. CAN YOU HELP FILL IN THOSE BLANKS FOR US ABOUT WHETHER WE NEED TO TEMPER OUR EXPECTATIONS IN THIS REGARD?

The caption changes to "Marzyeh Ghassemi, @MarzyehGhassemi."
Then, it changes again to "Limitations."

Marzyeh says SO I DO THINK THAT WE NEED TO TEMPER OUR EXPECTATIONS. AND WHAT I MEAN IS, WHEN YOU DESIGN THESE HIGH CAPACITY MODELS, THESE DEEP NEURAL NETWORKS, WHETHER THEY'RE WORKING ON IMAGES OR ON TEXT OR TABULAR DATA, THEY'RE OFTEN VERY FRAGILE. WHAT I MEAN IS THEY UNDERSTAND THE EXAMPLE SITES THAT THEY'VE BEEN SHOWN. AND IN MANY OTHER DOMAINS, WE HAVE BILLIONS OF IMAGES OF DOGS AND CATS, FOR EXAMPLE, FROM DIFFERENT COUNTRIES AND DIFFERENT SETTINGS, IN SNOW, INDOOR LIGHTING, OUTDOOR LIGHTING, DIFFERENT POSES. AND THESE MODELS CAN TUNE ALL OF THEIR PARAMETERS TO UNDERSTAND THIS IS GENERALLY WHAT A DOG LOOKS LIKE IN MANY DIFFERENT SETTINGS. THAT'S NOT THE SPACE THAT WE OPERATE IN IN HEALTH CARE. WE DON'T HAVE A BUNCH OF HEALTHY CHEST X-RAYS. WE DON'T HAVE A BUNCH OF HEALTHY HUMAN DATA. AND SO WHEN WE'RE SHOWING MACHINE LEARNING MODELS, NOT JUST SICK PEOPLE BUT OFTEN THE MOST DATA FROM THE VERY SICKEST PEOPLE, THEY'RE GOING TO BECOME SLIGHTLY MORE FRAGILE THAN WHAT WE'VE LEARNED TO EXPECT THESE SYSTEMS IN OTHER SETTINGS CAN BE. SO WE NEED TO BE VERY CAUTIOUS ABOUT AUDITING MODELS. AND WHEN WE DESIGN THEM, MAKING SURE THAT THERE'S ROBUSTNESS BAKED IN.

Steve says SO THIS IS A LIMITATION YOU HAVE OBSERVED IN TERMS OF AI'S CAPABILITIES. DOES IT GIVE YOU PAUSE AS TO HOW SUCCESSFUL YOU THINK YOUR ENDEAVOURS CAN BE BY THE END OF THE DAY?

The caption changes to "Marzyeh Ghassemi. University of Toronto."

Marzyeh says I THINK THAT, AS A RESEARCHER, I HAVE SO MUCH OPTIMISM ABOUT WHERE AI AND MACHINE LEARNING CAN TAKE MEDICINE. BUT THAT'S ON THE RESEARCH SIDE. I DON'T THINK THAT MANY OF THE TECHNOLOGIES WE DEVELOP ARE READY FOR DEPLOYMENT YET BECAUSE DEPLOYMENT HAS OTHER CONSIDERATIONS. HUMANS ARE USING THESE SYSTEMS ON OTHER HUMANS. SO I THINK YOU HAVE TO BE SIGNIFICANTLY MORE CAREFUL WHEN YOU GO FROM A RESEARCH ANALYSIS TO A REAL DEPLOYMENT IN A HUMAN SETTING.

Steve says NAHEED, I WONDER IF I COULD FOLLOW UP IN THIS REGARD. I KNOW LOTS OF PEOPLE WATCHING THIS RIGHT NOW ARE WONDERING WHY SO MANY OF THE WORLD'S TRULY, WE THOUGHT, WONDERFUL INSTITUTIONS, WERE UNABLE TO SEE THIS COVID-19 TRAIN COMING DOWN THE TRACKS MORE THAN A YEAR AGO, AND IT KIND OF CAUGHT US OFF-GUARD. DO YOU ANTICIPATE THAT THE KIND OF WORK THAT YOU'RE INVOLVED IN RIGHT NOW MIGHT BE ABLE TO HELP US PREDICT SEEING THAT TRAIN COMING DOWN THE TRACKS AHEAD OF TIME SO WE DON'T GET CAUGHT SO FLAT-FOOTED NEXT TIME?

The caption changes to "Naheed Kurji, @NaheedKurji."

Naheed says YEAH, SO THERE ARE TWO COMMENTS I'D LIKE TO MAKE. ONE, I FULLY AGREE WITH EVERYTHING THAT WAS STATED BEFORE AND IT'S A VERY IMPORTANT LIMITATION OF AI, THAT IT'S NOT THE SILVER BULLET, THERE'S NO ONE-SIZE-FITS-ALL SOLUTION. IN TERMS OF YOUR SPECIFIC QUESTION, WE AT CYCLICA ARE NOT IN THE BUSINESS OF PREDICTING THE NEXT OUTCOME. THERE IS SO MUCH WORK BEING DONE BY GOVERNMENT INSTITUTIONS, BY NON-PROFIT ORGANIZATIONS, AS WELL AS BY LIFE SCIENCE TECHNOLOGY-DRIVEN COMPANIES TO START PREDICTING THE NEXT INFECTIOUS DISEASE OUTCOMES. I'D SAY MORE THAN JUST THE PREDICTION, IT'S HOW DO WE GET AHEAD OF IT AND MINIMIZE THE EFFECT THAT IT WILL INEVITABLY HAVE. THERE'S GOING TO BE ANOTHER INFECTIOUS DISEASE. THERE IS SOME WONDERFUL WORK BEING DONE BY GOOGLE'S DEEP MIND IN USING THEIR AI TECHNIQUES TO UNDERSTAND THE BIOLOGICAL DRIVERS OF DISEASE, SUCH THAT WHEN NEW INFECTIOUS DISEASE, WHERE THERE'S NOT AN AMPLE AMOUNT OF DATA ON THAT DISEASE, LIKE COVID EARLY IN 2020, THAT WE CAN GET AHEAD OF IT AND START SPARKING A NEW WAVE OF INNOVATION TO STYMIE ITS IMPACT SO IT DOESN'T SHUT DOWN THE GLOBAL ECONOMY. SO IN TERMS OF WHAT CYCLICA IS DOING IN PREDICTING, THAT'S NOT OUR BUSINESS, BUT THERE'S SO MUCH HAPPENING IN THE WORLD ON THAT. OUR BUSINESS IS ALL-AROUND, WHEN IT DOES EXIST, WHAT CAN WE DO TO CREATE MEDICINES FASTER OR GET THOSE TO PATIENTS IN A MUCH MORE EXPEDITIOUS TIME FRAME.

Steve says MARZYEH, MAYBE I CAN GET YOU ON THAT TOO. HOW MUCH OF YOUR MISSION DO YOU SEE AS BEING MORE PREDICTIVE IN HOW YOU HANDLE FUTURE PANDEMICS?

Marzyeh says A LOT OF WHAT MY RESEARCH GROUP IS DOING RIGHT NOW IS TRYING TO UNDERSTAND WHEN WE CAN CATEGORIZE DATA OF BEING OUT OF DISTRIBUTION, WHICH IS A TECHNICAL TERM MEANING, IT DOESN'T LOOK LIKE THE DATA WE'VE BEEN SEEING BEFORE. I THINK THE EVENTUAL RESULT OF THAT KIND OF RESEARCH COULD BE IDENTIFYING WHEN WE'RE SEEING PATIENTS THAT SEEM LIKE THEY'RE NOT FROM THE DISTRIBUTIONS WE'RE EXPECTING. THIS ISN'T A NORMAL FLU SEASON, FOR EXAMPLE. BUT IN ORDER TO DO THAT, WE WOULD NEED A WELL-ALIGNED SET OF INTERNATIONAL DATA, RIGHT? BECAUSE AS THIS PANDEMIC SPREAD, WE WOULDN'T HAVE SEEN ANYTHING OUT OF DISTRIBUTION IN NORTH AMERICA, FOR EXAMPLE, UNTIL MUCH LATER ON. AND SO THAT WOULD REQUIRE COLLABORATION SO THAT WE CAN HAVE ALERTS THAT TELL US THAT THIS IS COMING DOWN THE LINE FROM ANOTHER COUNTRY.

Steve says NAHEED, YOU LOOKED LIKE YOU WANTED TO MAKE AN ADDITIONAL COMMENT ON THAT.

Naheed says NO, THAT'S ALL VERY WELL-STATED. WHEN I THINK ABOUT THE LIMITATIONS OF AI, IT COMES DOWN TO THE RELIANCE ON DATA. SO AI IS GLAMORIZED, IT'S GLORIFIED, PEOPLE ARE VERY EXCITED ABOUT IT. BUT THE REALITY IS, WHEN THE DATA IS NOT OF HIGH QUALITY, WHEN THERE'S NOT A LOT OF IT, AND WHEN IT'S NOT RELEVANT TO THE SPECIFIC RESEARCH PROBLEM, THE LIMITATIONS OF AI BECOME QUITE CLEAR. YOU KNOW, COMPUTERS ARE VERY FAST, AND THIS IS RELEVANT THROUGH MILLIONS AND BILLIONS OF DATA POINTS. THEY'RE ALSO LACKING INTUITION. SO THAT COMES BACK TO MY PREVIOUS POINT, THAT AI IS NOT THE SILVER BULLET, IT SIMPLY CAN'T BE, AND IT IS BEST WHEN IT IS AUGMENTED, ESPECIALLY IN THE SPACE THAT I THINK ABOUT FREQUENTLY, WHICH IS HEALTH CARE, WHEN IT AUGMENTS A HEALTH CARE PRACTITIONER, WHETHER IT BE A DOCTOR ON THE FRONT LINE, WHETHER IT BE A RESEARCHER THINKING ABOUT NEW MEDICINE DISCOVERY OR A RESEARCHER TRYING TO UNDERSTAND THE BIOLOGY, AI IS BEST WHEN IT AUGMENTS THOSE PEOPLE. IT'S NOT GOING TO REPLACE AND CERTAINLY NOT IN CONTEXT WHEN THERE'S A LACK OF DATA ACROSS THE THREE PILLARS THAT I JUST MENTIONED.

The caption changes to "Subscribe to The Agenda Podcast: tvo.org/theagenda."

Steve says TO THAT END, I ACTUALLY HAVE SOMETHING FROM WHAT THEY USED TO CALL THE BRITISH MEDICAL JOURNAL, THEY'RE SORT OF FACETIOUSLY ASKING THE QUESTION WHAT IS IT FOR ARTIFICIAL INTELLIGENCE OR AUGMENTING... WE'LL PUT THIS QUOTE UP AND COME BACK AND CHAT.

A quote appears on screen, under the title "Is A.I. exacerbating health inequity during COVID?" The quote reads "Among the most damaging characteristics of the COVID-10 pandemic has been its disproportionate effect of disadvantaged communities. As the outbreak has spread globally, factors such as systemic racism, marginalisation, and structural inequality have created path dependencies that have led to poor health outcomes... Artificial Intelligence (A.I.) technologies -quantitative models that make statistical inferences from large datasets- are an important part of the health informatics toolkit used to fight contagious disease. A.I. is well known, however, to be susceptible to algorithmic biases that can entrench and augment existing inequality. Uncritically deploying A.I. in the fight against COVID-19 thus risks amplifying the pandemic's adverse effects on vulnerable groups, exacerbating health inequity."
Quoted from David Leslie et al., BMJ. March 16, 2021.

Steve says OKAY. THAT'S A PRETTY STRONG CHARGE, BUT MARZYEH, LET'S GET SOME DISCUSSION GOING ON THAT. WHAT DO YOU THINK?

The caption changes to "The role of bias."

Marzyeh says I THINK IT'S AN ACCURATE CHARGE. I DON'T THINK THAT THERE'S ANY EXAGGERATION THERE. SO THE ISSUE IS THAT WHEN WE DO SCIENCE, THERE'S MANY STAGES IN THE PIPELINE WHERE BIAS CAN CREEP IN. SO STARTING FROM, WHAT KIND OF SCIENCE DO WE FUND? WHAT KIND OF DISEASES ARE WE ACTUALLY STUDYING? THERE ARE SEVERAL DISEASES THAT ARE JUST UNDERSTUDIED BECAUSE THERE'S NO FUNDING TO STUDY THOSE DISEASES BECAUSE THEY AFFECT MINORITIES OR MINORITY GROUPS. THE SECOND IS ONCE YOU'VE FUNDED A STUDY, MAYBE WE'RE STUDYING COVID. WHEN YOU RECRUIT PEOPLE, OFTEN YOU RECRUIT FROM THOSE WHO ARE CONVENIENT. BUT IF YOU'RE AN ACADEMIC RESEARCHER, WHAT'S CONVENIENT? MAYBE YOUR STUDENT POPULATION. THAT'S NOT REALLY A REPRESENTATIVE GROUP. IT'S NOT OFTEN DIVERSE. AND THEN EVEN IF YOU HAVE A DIVERSE SORT OF REPRESENTATIVE GROUP, WE KNOW THAT BIAS CAN CREEP IN IN THE WAY WE DEFINE OUTCOMES, IN THE WAY THAT WE CODE ALGORITHMS TO LEARN ABOUT THE MEAN OF A DISTRIBUTION, RATHER THAN THE TAILS OR THE MEDIAN. AND THEN WHEN YOU DEPLOY SOMETHING, OFTEN THERE'S OTHER CONSIDERATIONS LIKE, IF AN ALGORITHM AGREES WITH WHAT MY OWN INTUITION IS, MAYBE I HAVE MY OWN UNCONSCIOUS BIASES, I'LL LISTEN TO IT. BUT IF IT CONTRADICTS SOMETHING I THINK, I'M NOT GOING TO LISTEN TO IT. SO IT'S ONLY GOING TO FURTHER ENTRENCH MY NEGATIVE BIASES WHEN IT AGREES WITH ME OR IT WON'T LISTEN WHEN IT TRIES TO CONTRADICT THE NEGATIVE BIASES I HAVE. THERE'S A LOT OF DANGERS TO NAIVELY DEPLOYED AI. I GET PITCHED A LOT OF AI SOLUTIONS FOR PANDEMIC PROBLEMS LIKE, WHY DON'T WE BUILD A MACHINE LEARNING SYSTEM TO DETECT FACE MASKS IN PUBLIC AND FIGURE OUT WHO IS WEARING AND WHO IS NOT? MY COMMENT IS THAT SEEMS LIKE A PUBLIC HEALTH PROBLEM. WHY DON'T YOU JUST GIVE PEOPLE FACE MASKS AND CONTINUE WITH MESSAGING THAT'S SPECIFIC ABOUT THEIR USE?

Steve says IN MARZYEH'S COMMENTS ARE REALISTIC DANGERS, AS BMJ HAS POINTED OUT AS WELL, WHAT DOES EVERYBODY PLAN TO DO ABOUT THIS?

Naheed says SO IN ADDITION TO MY CAPACITY AS CO-FOUNDER AND PRESIDENT AND CEO OF CYCLICA, I'M ONE OF THE BOARD MEMBERS AND EXECUTIVE OFFICERS OF THE ALLIANCE FOR ARTIFICIAL INTELLIGENCE IN HEALTH CARE. INCREASINGLY OVER THE PAST 12 TO 18 MONTHS WHAT WE'VE BEEN TALKING A LOT ABOUT IS NOT ABOUT QUALITY, QUANTUM, AND RELEVANCE OF DATA BUT ALSO FAIRNESS OF DATA. AND WITHIN THE HEALTH CARE SPACE, WITH APPLICATIONS IN DISEASE DIAGNOSTICS, DRUG DISCOVERY AND DEVELOPMENT, DECISION SUPPORT, PATIENT CARE, DISEASE MANAGEMENT, AI IS FAST BECOMING THE INDUSTRY STANDARD. AND THE INTENT OF AI IS TO HELP HEALTH CARE PROVIDERS MAKE MORE OBJECTIVE DECISIONS AND PROVIDE MORE EFFECTIVE CARE. THAT IS IT. AND WHILE THERE ARE MANY REAL AND POTENTIALLY... YOU KNOW, BENEFITS OF USING AI IN GENERAL, AN IMPORTANT RISK IS [indiscernible] DECISION MAKING AND THAT'S BEDDED IN AI OUTPUTS AND A BY EXAMPLE OF THIS IS... THERE'S MANY EXAMPLES. YOU THINK ABOUT HR AND HOW, YOU KNOW, AI ALGORITHMS IS POINTED HR SPECIALISTS TO PEOPLE THAT ARE BIASED TOWARDS THE WAY IN WHICH THEY THINK AND THROUGH OTHER WAYS. IN THE HEALTH CARE SPACE, THIS IS DONE VIA SYSTEMIC REVIEW OF PAIN MANAGEMENT STUDIES. AN AUGUST 2020 ARTICLE IN A JOURNAL POINTED TO THE DISSEMINATION OF UNDERDEVELOPED AND POTENTIALLY BIASED MODELS IN RESPONSE TO NOVEL CORONAVIRUS. AND SO WHAT DOES THIS MEAN? YOU KNOW, THE MOST FREQUENT PROBLEMS ENCOUNTERED WERE UNDER REPRESENTATIVE DATA SAMPLES, MY LIKELIHOOD OF MODEL OVERFITTING IMPRECISE REPORTING OF STUDIED POPULATIONS AND THE CONSEQUENCE OF ALL OF THIS IS THERE IS A CERTAINLY MARGINALIZED GROUP OF PEOPLE WHO ARE SIMPLY NOT GETTING THE CARE AND THE TREATMENT THAT THEY REQUIRED.

Steve says IN WHICH CASE, MARZYEH, DO YOU KNOW WHETHER OR NOT YOUR HEALTH CARE COMMUNITY IS AWARE ENOUGH OF THIS TO MAKE SURE THAT IT IS NOT HANDCUFFED BY IT?

Marzyeh says I DON'T THINK THAT WE ARE AWARE ENOUGH YET. I THINK WE ARE RAISING THE BAR. SO I DO THINK THAT NOW THERE'S A MUCH MORE GENERAL AWARENESS OF THE POTENTIAL BIASES WHEN YOU DEPLOY A MODEL. I AM REALLY CONCERNED ABOUT THE POTENTIAL KNOCK-ON EFFECTS OF BIASES THAT WE DON'T CATCH, RIGHT? GETTING MORE DIVERSE DATA IS NOT A SIMPLE ONE-MONTH FIX, RIGHT? THIS IS SOMETHING THAT'S GOING TO REQUIRE A LOT OF TIME AND A LOT OF EFFORT. I ALSO THINK THAT IT'S NOT ALWAYS TRUE THAT OUR INTENT WHEN WE DEPLOY SOMETHING IN A HEALTH CARE SETTING IS TO IMPROVE PATIENT CARE. A LOT OF TIMES IT'S TO OPTIMIZE BILLING OR MAYBE TO GIVE DOCTORS MORE TIME OR MAYBE TO AUTOMATE A SPECIFIC HOSPITAL LOGISTICS PROCESS, AND SO THE PROBLEM IS, WE'RE DEALING WITH A HUMAN PROCESS, AND IT'S NOT ALWAYS CLEAR WHAT IS MEANT TO BE OPTIMIZED WHEN A HUMAN IS PERFORMING SOMETHING. AND SO IF WE SAY, WELL, LOOK AT WHAT THE HUMANS ARE DOING, JUST DO THAT, THAT SEEMS TO BE WORKING FOR THEM; DO IT FASTER, DO IT MORE EFFICIENTLY, DO IT WITHOUT A PERSON THAT I CAN SORT OF ARGUE WITH. NOW I HAVE AN ALGORITHM THAT HAS THE SHEEN OF OBJECTIVITY, BUT IT'S JUST EMBEDDING EXACTLY WHAT NAHEED HAS SAID, THE HUMAN BIASES THAT WERE ALREADY IN THE DATA THAT WE TRAINED IT WITH. AND SO I THINK WE JUST NEED TO BE CAREFUL WHEN WE LOOK AT THE DATA THAT WE HAVE, THE MODELS THAT WE'RE TRAINING, THE OUTPUTS THAT WE GET, AND THE WAYS THAT THOSE WILL BE USED, THAT WE'RE THOUGHTFUL ABOUT HOW WE'RE GOING TO DEPLOY THESE MODELS. AGAIN, I AM A HUGE OPTIMIST ABOUT THE PROGRESS THAT AI CAN MAKE. WE'RE CLEARLY DEALING WITH A MEDICAL SYSTEM THAT'S OVERLOADED. IT'S NOT AS IF THERE ISN'T OPPORTUNITIES. WE JUST NEED TO BE CAREFUL ABOUT HOW WE TAKE ADVANTAGE OF THIS KIND OF TECHNOLOGY.

Steve says CAN I JUST UNDERSTAND, MARZYEH, WHAT THAT MEANS WHEN YOU SAY WE'VE GOT TO BE CAREFUL ABOUT IT. HOW DO YOU IN FACT GO ABOUT MITIGATING THE BIAS IN AI MODELS THAT YOU'RE CLEARLY CONCERNED ABOUT?

Marzyeh says WE DEFINITELY NEED MORE DATA. WE DEFINITELY NEED TO DEVELOP BETTER TECHNIQUES TO DE-BIAS MODELS WHEN THE DATA ITSELF IS BIASED. AND FINALLY, I'M A HUGE FAN OF REGULATORY BODIES THAT ARE TRYING TO UNDERSTAND HOW BEST TO EVALUATE AI. AND SO THE FDA HAS A CLEARANCE PROGRAM FOR HOW TO EVALUATE MEDICAL AI BEFORE IT CAN BE CONSIDERED CLEARED AND OKAY TO BE IN A NEW SETTING. I THINK THAT WE SHOULD HAVE VERY ROBUST DEFINITIONS, AS DIFFERENT COUNTRIES WHO HAVE DIFFERENT POPULATIONS, OF WHAT WE THINK COUNTS AS BEING A SAFE ALGORITHM FOR DEPLOYMENT, BECAUSE THAT DOESN'T JUST HOLD RESEARCHERS TO THE SAME BAR THAT WE WOULD LIKE SOCIETY TO BE HELD TO, BUT ALSO INCENTIVIZES COMPANIES TO INVEST IN RESEARCH THAT WILL CLEAR A SPECIFIC ROBUSTNESS BAR.

Steve says OKAY. NAHEED, LET'S OPEN UP OUR DISCUSSION A LITTLE BIT FURTHER HERE AND TALK ABOUT THE CURRENT LANDSCAPE FOR AI IN CANADA. HOW ARE WE DOING VIS-A-VIS THE REST OF THE WORLD?

The caption changes to "Future of Artificial Intelligence in Canada."

Naheed says SO, YOU KNOW, CONTROVERSIALLY, CANADA HAS BEEN EXCEPTIONALLY GOOD FOR DECADES AT RESEARCH AND INVENTION. IT HISTORICALLY HAS NOT BEEN AS GOOD AT THE INNOVATION AND COMMERCIALIZATION OF THAT RESEARCH. I'D SAY THAT HAS CHANGED OVER THE PAST FIVE, SEVEN, TEN YEARS, AND WE'RE ON A STEADY STREAM OF... WHAT WE'RE SEEING IS LEGITIMATE TECHNOLOGIES BEING COMMERCIALIZED AND HAVING GLOBAL CONSEQUENCES IN A POSITIVE WAY. AI IS HAVING, I'D SAY, THE VECTORS OF INFLUENCE OF AI GLOBALLY ARE POINTING TO CANADA. THE BELTWAY BETWEEN THE UNIVERSITY OF WATERLOO ALL THE WAY UP TO MONTREAL, WORK BEING DONE AT THE UNIVERSITY OF ALBERTA, AND ALL THE WAY DOWN IN VANCOUVER. AND WHEN YOU LOOK AT THE RESEARCH THAT'S HAPPENING, MUCH OF THAT IS THEN BEING COMMERCIALIZED IN ENTITIES THAT ARE HAVING GLOBAL IMPACT. LET'S THINK ABOUT WHAT THE FOUNDATION, THE SFI FOUNDATION HAS FUNDED OVER THE PAST 12 TO 18 MONTHS IN COVID. THERE ARE A NUMBER OF BIG DATA COMPANIES THAT ARE DEPLOYING NEW TECHNOLOGIES TO EITHER UNDERSTAND THE DRIVERS OF DISEASE AS WELL AS ON THE DIAGNOSTICS AS WELL AS THE DRUG DISCOVERY. AND AS I MENTIONED, CIFAR HAS BEEN AT THE FRONT OF CENTRE OF THAT FUNDING, A NUMBER OF PROGRAMS, MANY OF WHICH ARE NOT JUST ON THE RESEARCH SIDE, YES, THERE'S A LOT, BUT MANY ARE IMPORTANT ON THE COMMERCIALIZATION SIDE. THE LAST THING I WILL SAY THERE ARE A NUMBER OF COMPANIES ONE COULD MENTION, BUT THOSE COMPANIES ARE NOW ATTRACTING GLOBAL INVESTMENT CAPITAL. AND INVESTMENT CAPITAL IS CRITICALLY IMPORTANT BECAUSE THAT SPARKS MORE INNOVATION. AND OVER THE PAST 12 TO 18 MONTHS WITH THE SPOTLIGHT OF THE WORLD ON HEALTH CARE AND ON AI AND THE IMPACT THAT IT CAN HAVE IN HEALTH CARE, WE'RE SEEING AN INCREASED FLOW OF CAPITAL TO SUPPORT THIS INNOVATION STRATEGY. SO I'M VERY BULLISH, THE CURRENT LANDSCAPE, AS WELL AS THE FUTURE LANDSCAPE, PRIVATE, PUBLIC, AS WELL AS GLOBAL INSTITUTIONS CONTINUE TO SUPPORT INNOVATION IN CANADA.

The caption changes to "Watch us anytime: tvo.org, Twitter: @theagenda, Facebook Live, YouTube."

Steve says MARZYEH, CAN I GET YOU ON THAT? WHAT DO YOU THINK MAKES CANADA PARTICULARLY ATTRACTIVE FOR AI AT THE MOMENT?

Marzyeh says THERE'S A JOKE IN THE MACHINE LEARNING COMMUNITY THAT THERE'S A CANADIAN MAFIA IN MACHINE LEARNING. AND IT'S SORT OF BASED IN REALITY THAT WE HAVE SOME OF THE MOST IMPRESSIVE RESEARCHERS IN DEEP LEARNING SPECIFICALLY IN THE WORLD IN CANADA. THERE'S NO OTHER PLACE THAT HAS SORT OF THIS LARGE SET OF FOUNDATIONAL DEEP LEARNING RESEARCHERS AND THAT'S ATTRACTED A LOT OF YOUNG TALENT IN THE RESEARCH SPACE. SO I THINK IN TERMS OF AI, WE ARE VERY WELL-PLACED TO DO INNOVATION. I THINK IN TERMS OF AI AND HEALTH, WE NEED TO SERIOUSLY RAMP UP THE MODERNIZATION OF OUR HEALTH CARE SYSTEM. WE'RE IN A NATIONAL HEALTH SYSTEM THAT DOESN'T SHARE DATA BETWEEN PROVINCES. WE DON'T CURRENTLY RECORD RATES IN OUR ELECTRONIC HEALTH CARE RECORDS SYSTEM. WE CAN'T CHECK WHETHER MORE BLACK CANADIANS ARE DYING THAN WHITE CANADIANS. SO WE NEED TO RAMP UP THE HEALTH SIDE OF WHAT WE HAVE BECAUSE THE AI RESOURCES ARE THERE.

Steve says IS THERE, IN YOUR JUDGMENT, AN INEVITABLE CLASH BETWEEN THE ACADEMIC SIDE OF THIS AND THE ENTREPRENEURIAL SIDE OF THIS?

Marzyeh says NOT NECESSARILY. I THINK A LOT OF TIMES, SPECIFICALLY WITH MACHINE LEARNING, THESE COMPANIES COME OUT OF SOMEBODY'S PhD. SO YOU HAVE A LOT OF STARTUP'S THAT ARE TRYING TO TAKE TECHNOLOGY THAT WAS PUSHED TO SORT OF THE FURTHEST IT COULD BE IN A RESEARCH SETTING AND THEN BRING IN THE PRACTICAL CONSIDERATIONS THAT ARE NECESSARY IN A DEPLOYMENT, AND SO THEY'RE GOING TO DO THAT IN A MORE STABLE CORPORATE ENVIRONMENT. AND SO I THINK THAT THERE'S A NATURAL SYNERGY THAT HAPPENS, AND WE'RE SEEING A LOT OF THAT IN TORONTO NOW AND ACROSS CANADA.

Steve says NAHEED, I'LL GIVE YOU THE LAST MINUTE ON THAT: NAVIGATING BETWEEN THE ACADEMIC AND THE ENTREPRENEURIAL SIDE. HOW IS IT GOING?

Naheed says INCREASINGLY SEEING MORE COLLABORATION, A PROPENSITY TO HAVE RESEARCH AND BRING IT COMMERCIALLY SO IT CAN HAVE THAT IMPACT. WHEN WE STARTED CYCLICA ABOUT SEVEN OR EIGHT YEARS AGO, MY EXPERIENCE WAS THAT WASN'T NECESSARILY THE ATTITUDE. OVER THE PAST NUMBER OF YEARS, AND IMPORTANT OVER THE PAST 12 TO 18 MONTHS, THAT ATTITUDE HAS CHANGED. AND YOU CAN WALK THE HALLWAYS OF LEADING ACADEMIC INSTITUTIONS IN CANADA, WHETHER IT BE IN MONTREAL, YOU KNOW, AT THE UNIVERSITY OF MONTREAL, McGILL, UNIVERSITY OF TORONTO WITH THE PROMINENT AND GLOBALLY IMPACTFUL VECTOR INSTITUTE, AS WELL AS THE UNIVERSITY OF ALBERTA, AND YOU CAN WALK THOSE HALLS AND PEOPLE AREN'T JUST TALKING ABOUT PUBLISHING THE NEXT PAPER AND TALKING ABOUT WHAT'S THE IMPACT FOR A SILOED RESEARCH PRODUCT, BUT NOW WHAT CAN WE DO ABOUT THAT AND WHAT CAN WE DO TO BUILD TEAMS AT SCALE? THAT LEVEL OF CONVERSATION TO ME IS FAIRLY UNIQUE. IT GIVES ME CONFIDENCE OF CONTINUING TO BUILD OUT CYCLICA IN CANADA. WE HAD THE OPPORTUNITY OF GOING TO THE U.S. MANY YEARS AGO, RAISING MUCH MORE MONEY. WE SAID WE'RE GOING TO INVEST IN CANADA. WE BELIEVE THE BEST TALENT AT THE INTERSECTION OF SCIENCE TECHNOLOGY IS HERE. AND WE'VE GONE FROM A FOUR-PERSON COMPANY TO NOW CLOSE TO A 60-PERSON COMPANY WITH A VAST MAJORITY OF THOSE BEING CANADIAN BASED AND EDUCATED AND WE'RE SEEING THE IMPACT OF THE INNOVATION AND THE COMMERCIALIZATION VERY STRONGLY CENTERED ON THE CANADIAN INNOVATION STRATEGY.

The caption changes to "Subscribe to The Agenda Podcast: tvo.org/theagenda."

Steve says WELL, IT'S SO NICE TO SEE THE ACADEMIC AND THE ENTREPRENEUR GETTING ALONG SO WELL HERE ON TVO TONIGHT. AND MARZYEH, I GUESS WE OWE A THANK YOU TO YOUR CO-HOST SITTING BESIDE YOU OUT OF CAMERA RANGE THERE FOR ALLOWING YOU TO GET THROUGH THE INTERVIEW PRETTY WELL UNSCATHED, SO WELL DONE ON THAT.

The caption changes to "Producer: Preeti Bhuyan, @preetibhuyan."

Marzyeh says SHE IS BEING REMARKABLY QUIET, YES.

Steve says WELL DONE ON THAT. MARZYEH GHASSEMI, ASSISTANT PROFESSOR AT UNIVERSITY OF TORONTO, NAHEED KURJI, FROM CYCLICA, IT'S GOOD OF BOTH OF YOU TO JOIN US ON TVO TONIGHT. THANK YOU.

Marzyeh says THANK YOU SO MUCH. TAKE CARE.

Watch: Can AI Combat COVID-19?