 All right now we're into our 15-minute presentations and we have Dajun Wang who's going to talk about understanding career hot streets All right. Okay, right. Hey, good morning everybody. It's a real pleasure to be here I want to tell you a little bit about a one journey we've been taking which is trying to understand what are the unfoldings What are patterns governing an unfolding career? And so this is like started I would say now looking back a decade ago Where with a number of colleagues that become fixated on this what I call the hope project Which is motivated by a set of very simple questions one is you know as a career unfolds When do you do your greatest work in your career and when you produce a higher impact work following a major? Breakthrough in other words is there a hope left or are we dead scientifically speaking? so We've been starting this for a decade now and I feel like we're been chasing the whole picture We still don't have it so every time we feel like we understand something and there are new questions that emerge So I still don't have the whole picture But we did understand a couple of patterns and so I want to share with you in particular one of them Which is the idea of a health streak and try to unpack of this idea So this is go back to five years ago My PhD student Lou Liu and I along with other collaborators We reported this a phenomena that we discovered by analyzing careers of artists over the past five hundred years direct film directors over the past Hundred years and and scientists from different disciplines over the past 50 years so analyzing all this different careers As different that they may seem we uncover an interesting Intriguing fact which is they all share a common feature. That's what we call hot streaks These are bursts of high-impact works occur together in close succession. Okay, so once we analyze this careers We realize there are a couple of patterns with the health streaks if this is forwarding, okay That is first we realize how streaks are rather ubiquitous and you know Within each domain about 90% of people have at least a health streak second is usually unique most people have just one and Then we also realize that how strict actually occurs in terms of within the sequence of the work you produce along your career It occurs remarkably randomly within the life cycle. Okay, so with equal probability This may occur on-site with your very first work to roll the last work or somewhere in the middle So this adds important nuances to the traditional wisdom that you do your best work in your mid-career What we're seeing here is that you're really just producing a lot more in your mid-career It doesn't mean the age and creativity are intertwined. We see a remarkable constant rate of being hit states your probability of Creating a hit paper or hit work stays remarkably constant Throughout the career and we also see that it does not last forever So for most people this lasts about four to five year period And then you fall back to where you were before and appears unassociated with productivity change In other words, you don't really produce more than we would expect back to otherwise within this period It's just what you produce during this period seems substantially better than everything else Yeah, so when the paper was first published, you know, and and I think this is one of those things of like reading more questions You know, I when the paper was published So I feel like it was deeply unsatisfying because we just reported at this phenomena We had no idea what are the identifiable regularities underneath the beginning of a health streak And that's why I wrote this kind of a reason new questions to me is that, you know, is there any? Identifiable regularity underneath the beginning of a health streak and and so just to be clear the idea of a health streak Obviously is quite important Especially think about science The projection of future impact is very important to a range of decisions that may make Including hiring promotion, tenure and research support So so if you actually ignore health streak and the singular nature, you know You may actually systematically over underestimate an individual's true potential, right? So to me these are essential questions because if we can understand this question It will allow us to answer a range of new questions For example, can we anticipate the start of a house trick or the end of it Can we create an environment to facilitate the outside of how strict and who extended when the emerges and for someone who has had a How strict can we treat it as an indication of that person's potential and help them realize that potential again? So so I read this question shortly after the publication of first paper. I was optimistic I'm gonna find the answer and next year. I'm gonna publish a paper So what happened the two years after that was repeated at times of failure So we tried one hypothesis after another we just couldn't find one that will explain what we're observing. So so then I was Presenting the other day at an IH symposium with all the program director just sitting there About this finding and a program a nice program director came to me. That's when can't you see it? This is so simple. Did you say house week is four to five year period? So we funded it, you know, that's like, you know, it's so simple like why that happens because our are ones We funded them and I said, okay, it does not explain artists, but you know, it's actually a Testable prediction. So why don't I test it? So I just get all your investigators See when they got their first hour one See where their house week is and see where the hour one comes is doesn't come before the house week or it comes somewhere else So we analyzed that systematically was we see is that your hour one funding didn't come before the house week It come exactly when the house week about to end. Okay, so Now you realize this actually makes a lot of sense because when you publish one nature paper after another That's where you start getting a lot of money. So it is consistent with the emphasize on preliminary work It is a protecting a nice investment But at the same time it does raise question in my mind that are you finding streak or you're finding retirements So this is like a simple illustration of sort of thinking on the idea of a house week is quite simple But how do we incorporate this into our daily? You know decision-making and calculus in general thinking about careers. It's actually easier said than that Okay, so so we've been searching for a lot of answers and then there's a epiphany moment for me When I was walking around the Van Gaal Museum in Amsterdam. I'm art armatures I don't know much about art But I was I know Van Gaal is a data point in my sample where his house streak happened in 1888 when he moved from Paris to South of France So that's where I read when I was looking around the this paintings. I realized. Oh, you know, even for someone who don't know art I can actually see some difference maybe before 1888 and afterwards So that's where I realized if we want to actually study what happens in the house week We have to go deep into what the character of the work is. Okay So then my students, you know Lou Lou and I along other Group members started to collect completely new data sets And so these are chasing not just the careers of artists and artwork they produce But also look at what these artworks are in terms of images of the artwork So we have 800,000 images for movies We look at the plots and the cost information of each movie and for publications is actually easier because we have the Publications and then you quickly realize we run into a problem is 800,000 images are gonna take someone a bit to actually go through So that's where we realize, you know, sort of if we we can actually develop this deep neural network by repurpose image a popular image recognition deep neural network and Connected to a fully connected layer to basically repurpose that to classify paintings And that's where you see actually you can achieve 97% accuracy But what we're interested is to open up the hood of these neural networks and look at different layers that give us Different levels of abstraction and say tell us what are you seeing that gives you this predictive accuracy? So we can do that systematically for all the paintings and similarly We can do that for plots in terms of world inviting or thinking about a cast of a movies in terms of no Inviting methods at the popular deep neural network methods and for papers We can also twist the references of different papers and see What are the topics of each paper as a individual careers unfold? Okay, so after we develop all this machinery now we can test the some hypothesis Out of many factors as being considered to affect career progression of success The idea of exploration exploitation has been commonly considered to think about influence this In the very broad range of literature so that lead us to ask, you know our career house weeks reflective Exploration or exploitation strategy So what we could do is just line everybody up as to where the house week occurs Look at the work before and afterwards and my majority entropy of the work Okay, and then so that we find three simple findings first What we see is that before the house week occurs your entropy is systematically larger than I expected So individuals tend to diversify the topics they work on so this is a before the house week Those are consistent with exploration strategy of trying different topics and following the outside of house week We see that entropy becomes systematically smaller than I expected so you become more focused on what you work on So this is consistent with exploring Exploitation strategy and third is that we see despite the differences across three domains we study and the methodology is we use to Study them this idea of exploration Exploitation and house week seems to consistant across all these domains Importantly what we realize is that actually neither exploration nor exploitation alone explains house week It's really this exploration followed by exploitation where the transition from exploration to exploitation Closely choices where your house week occurs. So this is sort of a plausible explanation for this is that Exploration is a risky strategy. So increases Your variance and the sample found some great ideas the subsequent Exploitation then allow you to focus build on knowledge and capabilities in that area and foster reputation related to that expertise So then the idea of exploration before Exploitation and serves to expand an individual's creative possibilities So this discovery of exploration follow exploitation then convince me to try to understand. Okay What about how do we start the first step? Okay, how do people explore? So that's the question of you know, how effectively we do researchers actually pivot and change their research directions And this is surprisingly understudied question And we also know that science and society face an evolving array of questions problems opportunities So researchers needs to constantly adjust their research stream and also the concept of Adaptability has also been considered a long-standing Being more essential to the survival and performance of firms economies organizations and society But worry little is known about how and how successfully Researchers may adapt to shifting demand. So in this paper then follow on paper with colleagues of my Ryan Hill Ian who is here a Caroline staying in Ben Jones We want to study the adaptability of scientists as inventors. So we presented two levels of analysis versus the general facts using millions of papers and patterns and see when people shift what happens and then second is application use case, which is a COVID-19 where in 2020 about four to five percent scientists shifted their career They're trajectory to engage with COVID-19. So over does what we discovered is a very pessimistic picture In contrast to earlier picture may be more optimistic. So this is to me very pessimistic Which is this idea of a pivot penalty So what we we develop a measure to quantify how far you are venturing out of your core area of competency And what we see is that the further you go the impact of the work just plummets Okay, systematically declines and so this is what we call a pivot penalty as you pivot It seems to be a great penalty to the impact of the work And we look at this across our scientists and inventors it holds for scientists and inventors It holds for different fields of science different patenting domains and also over time it gets worse Okay, so we look at it over the past five decades We see every decade that penalty grows worse and worse harder and harder steeper and steeper So but what about COVID-19? We know COVID-19 had a lot of demands for the research. So it's a high impact premium so when we apply this COVID-19 we see while you're right, there is an impact premium, but the Penalty is so steep that they sharply the offsets that impact premium in other words If you're already working on COVID-19 related research, then you're gonna be great You're gonna do very well, but if you're actually travel worry far to engage with this kind of research Then there's this deep penalty to your work. So what is tells us is on the individual level? There is a deep domain emphasize on deep domain expertise It's the individual become more and more specialized and the science policy level This is also tell us a limited opportunity to shift our resources and personnel's because of the scientific workforce appears to be Rather not worry adaptive. So just concluding remarks. I know you are standing up. So my time is almost up So I want to use this as an illustration of with the emergence of new data sets and tools like AI there are now a enormous opportunity to better understand and improve science and You know and I think science of science and meta science in general is producing new discoveries that can Reshape and accelerate the scientific and innovation ecosystem worldwide. I think policy wise We also know that there's now enormous opportunity to do more and to do better and the stakes are extremely high I think what motivates my work in thinking about how do we improve science is that if we can make R&D even 5% more Efficient the return to society will be enormous in terms of improving higher standards of living Longer and healthy life. Thank you. Yeah We have time for one very short question very short answer, please Thank you so much for your presentation Daniel of alley Michael Smith Health Research BC in Canada Question for you with respect to hot streaks. Did you look at the impact on the people that they mentor that are in their labs? I saw the social network with actors and directors for example, and you can imagine that it would then possibly propagate more hot streaks for the people that they mentor That's a great question see there's a new question that emerged that we haven't looked at I don't have the answer to that, you know And I think in general the month or month tea data sets by the way is just becoming available In the larger scale than previously it was available So I think a number of groups right now are trying to work on that data Maybe one crucial aspect is to try to understand how how strict properties. Yeah All right, the other one is you're gonna have to track a day John down at lunch and ask your question then all right Thank you very much. All right