video_id
stringlengths 11
11
| text
stringlengths 361
490
| start_second
int64 0
11.3k
| end_second
int64 18
11.3k
| url
stringlengths 48
52
| title
stringlengths 0
100
| thumbnail
stringlengths 0
52
|
|---|---|---|---|---|---|---|
McV4a6umbAY | except if either player throws, what is it? Scissors? Yes. If other player throws scissors, then the winner gets two points, and the loser loses two points. That's just to break the symmetry of the game. So, the next equilibrium in this game is to throw rock and paper with 40 percent probability, each, and scissors with 20 percent probability. Now, imagine that we're trying to | 2,441 | 2,464 | https://www.youtube.com/watch?v=McV4a6umbAY&t=2441s | AI for Imperfect-Information Games: Beating Top Humans in No-Limit Poker | |
McV4a6umbAY | do a depth-limited version, a depth-limited solving of this game as player one. So, we look one move ahead, and then we're going to substitute the Nash equilibrium value at each of those states, where instead of going to the end of the game. This is the depth-limited subgame. It's really easy to see that if we were to try to solve this depth limited subgame, | 2,464 | 2,485 | https://www.youtube.com/watch?v=McV4a6umbAY&t=2464s | AI for Imperfect-Information Games: Beating Top Humans in No-Limit Poker | |
McV4a6umbAY | there is no way that we're going to find the optimal strategy of 40 percent, 40 percent, 20 percent. Right? There's no just not enough information in this depth-limited subgame to find the Nash Equilibrium. Why is that? Well, it turns out the reason is because we are essentially assuming that beyond this decision point, player two is going to play the Nash equilibrium strategy. | 2,485 | 2,507 | https://www.youtube.com/watch?v=McV4a6umbAY&t=2485s | AI for Imperfect-Information Games: Beating Top Humans in No-Limit Poker | |
McV4a6umbAY | Right? That's where we got the 000 from, this is assuming that if we had chosen scissors, and player two plays the Nash equilibrium strategy beyond this point, this expected value is zero.. But in reality, player two strategy beyond the depth-limit depends on what our strategy is above the depth-limit. If we choose rock, 80 percent of the time, player two's strategy isn't going to | 2,507 | 2,527 | https://www.youtube.com/watch?v=McV4a6umbAY&t=2507s | AI for Imperfect-Information Games: Beating Top Humans in No-Limit Poker | |
McV4a6umbAY | be to play the Nash equilibrium, it's going to be choosing paper, 100 percent of the time. So, this is what the state values will look like or if we were to choose paper, 80 percent of time, then they would switch to always choosing scissors, and this is what the state values would look like. So, in an imperfect-information game, the values of the states of the depth-limit depend on what our policy is, | 2,527 | 2,550 | https://www.youtube.com/watch?v=McV4a6umbAY&t=2527s | AI for Imperfect-Information Games: Beating Top Humans in No-Limit Poker | |
McV4a6umbAY | in the earlier parts of the game. So, how do we deal with this? Well, one option is to just actually make the state values dependent on our policy, and say, "Okay. Well, the value of a state is a function of the description of that state, and our policy for the entire game." Well, that is theoretically correct to the promise that's extremely expensive, I mean absurdly expensive. | 2,550 | 2,573 | https://www.youtube.com/watch?v=McV4a6umbAY&t=2550s | AI for Imperfect-Information Games: Beating Top Humans in No-Limit Poker | |
McV4a6umbAY | The other option is something called, well, we actually metaphor about this AI called DeepStack. They condition the value of a state on the belief distribution of both players at that state. So, they said, okay. Well, at this decision point, I'm not going to condition on the strategy for the early part of the game, I'm going to look at all the different states that I might be in, | 2,573 | 2,596 | https://www.youtube.com/watch?v=McV4a6umbAY&t=2573s | AI for Imperfect-Information Games: Beating Top Humans in No-Limit Poker | |
McV4a6umbAY | and the probability that I believe I'm in each of those states, which in this case is 80 percent, 10 percent, 10 percent. Then, this game, it ends up being the same exact thing as just conditioning on the policy but, in general that's not the case. The problem is that this is still extremely expensive. So DeepStack for example, in Heads up No-limit Texas Hold'em, | 2,596 | 2,615 | https://www.youtube.com/watch?v=McV4a6umbAY&t=2596s | AI for Imperfect-Information Games: Beating Top Humans in No-Limit Poker | |
McV4a6umbAY | use 1.5 million core hours of computation, and could not be prior top AIs. The other problem is that the technique currently does not scale to larger games, basically games where you have more states in an information set. This, you can get by with this in Heads-up No-Limits Texas Hold'em, because it's only about, in any single decision point, every single information set, | 2,615 | 2,637 | https://www.youtube.com/watch?v=McV4a6umbAY&t=2615s | AI for Imperfect-Information Games: Beating Top Humans in No-Limit Poker | |
McV4a6umbAY | there's only about 1,000 different states in that information set. But in a game like Five Card Draw for example, there could be 5 billion or in a game like Stratego, that could be 10th to the 20th or something. So, this would not work in those larger games. So, we do instead, is this paper that we just had except it's NIPS for 2018 called a depth-limited solving, | 2,637 | 2,662 | https://www.youtube.com/watch?v=McV4a6umbAY&t=2637s | AI for Imperfect-Information Games: Beating Top Humans in No-Limit Poker | |
McV4a6umbAY | we use this in an AI, we created called Modicum, and let me walk you through the algorithm here. The idea is, instead of assuming that there is a single value at this depth-limit, we're actually going to let the opponent choose between multiple different values for these states. We create these different values in an iterative process. So, we start off by assuming | 2,662 | 2,684 | https://www.youtube.com/watch?v=McV4a6umbAY&t=2662s | AI for Imperfect-Information Games: Beating Top Humans in No-Limit Poker | |
McV4a6umbAY | that player two, beyond the depth-limit, is going to play the Nash equilibrium strategy that we precomputed. Then, we solve this depth-limited subgame, which in that case means, let say, we solve it, and we say our strategy is going to be one-third, one-third, one-third probability. Now, we're going to calculate a player two best response. So, we can say, okay. | 2,684 | 2,705 | https://www.youtube.com/watch?v=McV4a6umbAY&t=2684s | AI for Imperfect-Information Games: Beating Top Humans in No-Limit Poker | |
McV4a6umbAY | Well, if our strategy here is to play one-third, one-third, one-third, player two could exploit us by always choosing rock. Now, we add that best response to the set of strategies that player two can choose at the depth-limit. So, now, we're going to solve this depth-limited subgame again, and we're going to say, add this depth-limit, now player two can choose. | 2,705 | 2,727 | https://www.youtube.com/watch?v=McV4a6umbAY&t=2705s | AI for Imperfect-Information Games: Beating Top Humans in No-Limit Poker | |
McV4a6umbAY | They can either choose the Nash equilibrium strategy that we had before or they can choose this best response that we just calculated, which was always play rock. They make this decision because all of these states, sharing information set, they can't say, "Okay, well, in this state, I'm going to choose this policy, and this state I'm going choose this policy." | 2,727 | 2,745 | https://www.youtube.com/watch?v=McV4a6umbAY&t=2727s | AI for Imperfect-Information Games: Beating Top Humans in No-Limit Poker | |
McV4a6umbAY | They have to make the same decision at each of these different states that share an information set. So, we solve this depth-limited subgame again. Then, we again calculate a player two best response to that strategy, and then we add that strategy to the set of best responses that player two can choose at the depth-limit. We can repeat this process as many times as we want. | 2,745 | 2,766 | https://www.youtube.com/watch?v=McV4a6umbAY&t=2745s | AI for Imperfect-Information Games: Beating Top Humans in No-Limit Poker | |
McV4a6umbAY | Now, some details about this technique, it might seem like this is really expensive, and may not get good performance because we can't add like a million different strategies for player two to choose that the debt limit, but it turns out that because they are making this choice separately at each information set, they're essentially able to, even if we only give them a choice | 2,766 | 2,786 | https://www.youtube.com/watch?v=McV4a6umbAY&t=2766s | AI for Imperfect-Information Games: Beating Top Humans in No-Limit Poker | |
McV4a6umbAY | between 10 different strategies, beyond the depth limit, if they are making that choice individually at 100 different information sets, they're actually choosing between 10 to the 100 different strategies for the entire remainder of the game. So, it actually grows very quickly. The other thing for player one, I talked about what player two does, so player two is choosing | 2,786 | 2,809 | https://www.youtube.com/watch?v=McV4a6umbAY&t=2786s | AI for Imperfect-Information Games: Beating Top Humans in No-Limit Poker | |
McV4a6umbAY | between these different strategies that they could play for the remainder of the game, player one, we're going to assume is playing the approximate Nash equilibrium. Player one is us, we're going assume that weren't playing according to the approximate Nash equilibrium strategy for the remainder of the game. The set of player two's strategies is precomputed it's not determined in real time. | 2,809 | 2,828 | https://www.youtube.com/watch?v=McV4a6umbAY&t=2809s | AI for Imperfect-Information Games: Beating Top Humans in No-Limit Poker | |
McV4a6umbAY | They could do it in real time, it would be too expensive in practice. Okay. So, this is what performance looks like if we do this depth-limited solving thing. So, on the x-axis here, we have the number of values per leaf node that the opponent can choose between at the depth limit, and on the y-axis, we have exploitability measured in milligrams per game. This is a simplified version of No-Limit Access Hold'em, | 2,828 | 2,852 | https://www.youtube.com/watch?v=McV4a6umbAY&t=2828s | AI for Imperfect-Information Games: Beating Top Humans in No-Limit Poker | |
McV4a6umbAY | that only has two betting rounds instead of four. You can see, if we only assume, that if we assume that each state has a unique value, which is essentially the Nash equilibrium value like we would in a perfect information game, exploitability is extremely high. But as we add more strategies for the appointed to choose between at the depth limit, exploitability drops off very quickly. | 2,852 | 2,874 | https://www.youtube.com/watch?v=McV4a6umbAY&t=2852s | AI for Imperfect-Information Games: Beating Top Humans in No-Limit Poker | |
McV4a6umbAY | In fact, with 16 different choices, were essentially at Nash equilibrium. So, the beautiful thing about this. Yes. >> So why is that have one. [inaudible] blue will be the same.. >> No. Because actually blue, sorry, so for blue here, it's not doing any real-time reasoning, it's doing this like, if they had bet $60, I'm going to round that to $50. So, red is actually doing real-time reasoning, | 2,874 | 2,903 | https://www.youtube.com/watch?v=McV4a6umbAY&t=2874s | AI for Imperfect-Information Games: Beating Top Humans in No-Limit Poker | |
McV4a6umbAY | but it's assuming that each value has a well defined unique value. >> So, [inaudible] generation or frequency or something, so what [inaudible]. >> There are a lot of similarities between this and things like double oracle methods and things like that. Yes. All right. So, in terms of head-to-head performance, the really cool thing about this technique is that it allows us to make | 2,903 | 2,925 | https://www.youtube.com/watch?v=McV4a6umbAY&t=2903s | AI for Imperfect-Information Games: Beating Top Humans in No-Limit Poker | |
McV4a6umbAY | a really strong poker AI using very few resources. To give you some perspective, we had this bot called Tartanian8 that we made in 2016, which won the Annual Computer Poker Competition, which is competition among poker AIs. It used two million core hours of computation, 18 terabytes of memory, and there's no real-time reasoning. We have Slumbot which won the, | 2,925 | 2,945 | https://www.youtube.com/watch?v=McV4a6umbAY&t=2925s | AI for Imperfect-Information Games: Beating Top Humans in No-Limit Poker | |
McV4a6umbAY | that wasn't us, this is a different bot, that won the 2018 competition, used 250,000 core hours, two terabytes of memory, no real-time reasoning. Modicum which is the bot that uses its depth limited solving, uses just 700 core hours, 16 gigabytes of memory, plus real-time with a 4-core CPU in under 20 seconds per hand, and it beats both of those other bots. So, to put this in perspective | 2,945 | 2,969 | https://www.youtube.com/watch?v=McV4a6umbAY&t=2945s | AI for Imperfect-Information Games: Beating Top Humans in No-Limit Poker | |
McV4a6umbAY | even further, the broadest, which is the AI that we played against the top humans, used millions of core. I think it was something like two million or five million core hours, probably it's 20 terabytes of memory, and played in real-time using 1,000 cores. So, we're able to get what is essentially probably superhuman, we haven't actually tested against humans but I'm | 2,969 | 2,990 | https://www.youtube.com/watch?v=McV4a6umbAY&t=2969s | AI for Imperfect-Information Games: Beating Top Humans in No-Limit Poker | |
McV4a6umbAY | pretty sure this is a superhuman poker AI. We're able to get basically superhuman performance using the resources in a laptop. In fact, since I published this paper, I've just put another paper on archive where we figured out how to make this three times faster. So, it could probably run on a smartphone now. Yeah. >> Where is the [inaudible] you can just run. | 2,990 | 3,012 | https://www.youtube.com/watch?v=McV4a6umbAY&t=2990s | AI for Imperfect-Information Games: Beating Top Humans in No-Limit Poker | |
McV4a6umbAY | >> It turns out that there is a huge amount of variance in poker. >>Yes. >> Because we're doing real-time reasoning and we're taking 20 seconds per hand, the variance is massive. In fact, we actually, we train this using 700 core hours, it took us like a million core hours to actually compute all these results. So, this has been a problem in the entire field, | 3,012 | 3,031 | https://www.youtube.com/watch?v=McV4a6umbAY&t=3012s | AI for Imperfect-Information Games: Beating Top Humans in No-Limit Poker | |
McV4a6umbAY | that the variance is just absurd. So, this is a graph of head-to-head performance. Here we have the Libratus, which beats up humans. Here is Modicum which is the AI we just created that uses way fewer resources. Here are some other benchmark bots. A bot from 2016, a bot from 2014. Here is DeepStack, which is a bought from the University of Alberta, which actually has very low exploitability. | 3,031 | 3,054 | https://www.youtube.com/watch?v=McV4a6umbAY&t=3031s | AI for Imperfect-Information Games: Beating Top Humans in No-Limit Poker | |
McV4a6umbAY | But in terms of head-to-head performance didn't end up being that strong. It also uses this real-time reasoning as well. Though a different form of it. All right so the key takeaways, yes. >> You said that one doesn't have, it has low exploitability but it's not that strong? >> In terms of head-to-head performance it's not as strong. So, in terms of head-to-head performance it actually doesn't | 3,054 | 3,079 | https://www.youtube.com/watch?v=McV4a6umbAY&t=3054s | AI for Imperfect-Information Games: Beating Top Humans in No-Limit Poker | |
McV4a6umbAY | beat prior benchmark bots. >> Yes, I guess that's curious to me, because if you're not exploitable. Okay so-. >> When I say low exploitability, I mean just relative to the past bots. So, the exploitability is still, it could be extremely high, we actually don't know. We can't calculate exploitability exactly and heads up no limit takes it whole so we don't know. | 3,079 | 3,102 | https://www.youtube.com/watch?v=McV4a6umbAY&t=3079s | AI for Imperfect-Information Games: Beating Top Humans in No-Limit Poker | |
McV4a6umbAY | But it appears that it has lower exploitability compared to the absurdly high exploitability of the previous bots Yes. So, key takeaways. In real-time planning, you always have to consider how the opponent can adapt to changes in your policy. That is something that is really important in imperfect permission games. Perfect-information games you can mostly ignore that but not completely. | 3,102 | 3,128 | https://www.youtube.com/watch?v=McV4a6umbAY&t=3102s | AI for Imperfect-Information Games: Beating Top Humans in No-Limit Poker | |
McV4a6umbAY | Imperfect-information subgame cannot be solved in isolation. States and imperfect-information games do not have well-defined values. All right. So, I have done some other work that I did not discuss. One thing I did not discuss, well I guess I talked about it briefly is how we actually solve these games. We use an algorithm called counterfactual regret minimization, | 3,128 | 3,152 | https://www.youtube.com/watch?v=McV4a6umbAY&t=3128s | AI for Imperfect-Information Games: Beating Top Humans in No-Limit Poker | |
McV4a6umbAY | which has been around for about 10 years now. Works extremely well, even though the theoretical guarantee on convergence is only 1 over square root t. I just had a paper that I released, where I developed a new form of CFR which beats the prior state-of-the-art by a factor of three. So, I'm really excited about that, and that's going to be used in all the future research. | 3,152 | 3,174 | https://www.youtube.com/watch?v=McV4a6umbAY&t=3152s | AI for Imperfect-Information Games: Beating Top Humans in No-Limit Poker | |
McV4a6umbAY | I have some work on pruning in CFR. So, it turns out that CFR you end up exploring the entire game tree which is a big waste, because a lot of actions are suboptimal and you don't want to waste time coming up with what you should do if you play a really crappy poker hand, because in reality you just fold it right away. So, I have this pruning technique that provably reduces to | 3,174 | 3,195 | https://www.youtube.com/watch?v=McV4a6umbAY&t=3174s | AI for Imperfect-Information Games: Beating Top Humans in No-Limit Poker | |
McV4a6umbAY | computing and memory costs of running CFR asymptotically. In practice, it speeds things up by an order of magnitude of two. I also have a paper on determining the optimal actions in a continuous action space. So, one of the interesting things about no limits exist hold them is that you have this continuous action space where you can choose to bet any amount between $100 to $20,000. | 3,195 | 3,218 | https://www.youtube.com/watch?v=McV4a6umbAY&t=3195s | AI for Imperfect-Information Games: Beating Top Humans in No-Limit Poker | |
McV4a6umbAY | The truth is that, what we do right now is just domain-specific abstraction techniques, where we say okay, well, you probably just want to bet either half the pot or one times the pot or two times the pot, and it doesn't really matter if you are betting 0.6 times the pot or 0.5 times the pot. But that relies on domain knowledge that we know the optimal bet fractions are roughly in that range. | 3,218 | 3,242 | https://www.youtube.com/watch?v=McV4a6umbAY&t=3218s | AI for Imperfect-Information Games: Beating Top Humans in No-Limit Poker | |
McV4a6umbAY | So, it'll be nice to have an algorithm that doesn't rely on domain knowledge that can actually determine the bets to use without any human knowledge. So, that's where this 2014 paper does, and hopefully we'll have some follow-up work on that in the future. For future directions. So, I mentioned before, a lot of this work has been looking at poker as a domain, | 3,242 | 3,264 | https://www.youtube.com/watch?v=McV4a6umbAY&t=3242s | AI for Imperfect-Information Games: Beating Top Humans in No-Limit Poker | |
McV4a6umbAY | and that's because it takes a lot of infrastructure to actually build up to do experiments on large games. So, we have a lot of expertise that has been developed over the years on how to run these techniques efficiently on a game like poker. And if we wanted to test on another large game, it would take years to build up that expertise on how to do experiments on those other games. | 3,264 | 3,282 | https://www.youtube.com/watch?v=McV4a6umbAY&t=3264s | AI for Imperfect-Information Games: Beating Top Humans in No-Limit Poker | |
McV4a6umbAY | There's also no other really good, we can't really compare performance to other bots in other games, because there are no other games where there are bots that are competitive. But I would love to move beyond poker, and a few different directions with that. One is, we have these techniques for perfect information games like AlphaZero, and we have these techniques for | 3,282 | 3,303 | https://www.youtube.com/watch?v=McV4a6umbAY&t=3282s | AI for Imperfect-Information Games: Beating Top Humans in No-Limit Poker | |
McV4a6umbAY | imperfect-information games like poker, and it'll be nice to bridge the gap and find a single algorithm that works really well in all these different games. Another thing that I'm really interested in is going beyond two-player zero-sum games. So, I mentioned that if you try to move on to general-sum games it's a bunch of theoretical challenges the pop-up. So, right now, we don't know how to | 3,303 | 3,325 | https://www.youtube.com/watch?v=McV4a6umbAY&t=3303s | AI for Imperfect-Information Games: Beating Top Humans in No-Limit Poker | |
McV4a6umbAY | cope with those challenges. But dealing with general-sum games is really important because most real-world situations are general-sum and not not zero-sum, except for like maybe military interactions or security. So, in particular working out something like negotiation, I think is a really interesting line of research. In general moving things more towards real-world domains, | 3,325 | 3,346 | https://www.youtube.com/watch?v=McV4a6umbAY&t=3325s | AI for Imperfect-Information Games: Beating Top Humans in No-Limit Poker | |
McV4a6umbAY | I think we're at the point right now where we can actually start bringing these techniques into the real world, and I think that's going to be a really interesting line of research as well. All right. So, I'll stop there and I'll take some last minute questions thank you. >> Yes so, I guess, in the real world often you don't know the rules of the game, you don't know them in advance, [inaudible] situation. | 3,346 | 3,371 | https://www.youtube.com/watch?v=McV4a6umbAY&t=3346s | AI for Imperfect-Information Games: Beating Top Humans in No-Limit Poker | |
McV4a6umbAY | >> Right. >> But you do observe the outcomes of players. Have you thought about trying to, what works with race when you go into a situation where you don't know the rules of the game but you can observe the outcomes of players? >> That's a good question so yes. So, all of our work assumes that we have an accurate model of the world or the game that's being played. | 3,371 | 3,389 | https://www.youtube.com/watch?v=McV4a6umbAY&t=3371s | AI for Imperfect-Information Games: Beating Top Humans in No-Limit Poker | |
McV4a6umbAY | I think a lot of these techniques will carry over if you want to try to figure out the structure of the game as you go. There was actually a paper from another student at CMU recently on this problem. So, people are starting to look in this direction, I have not. But it's something that I think is very interesting. It is also a way harder problem, because to figure out | 3,389 | 3,409 | https://www.youtube.com/watch?v=McV4a6umbAY&t=3389s | AI for Imperfect-Information Games: Beating Top Humans in No-Limit Poker | |
McV4a6umbAY | what information your opponent knows or figuring out what they could know is a really difficult challenge. I'm not sure how to go about that. Yes. >> I think in many applications in something like reinforcement learning. So, right now I can imagine splitting the environment into the portion that which I can model that will be like Chess moves, and then I can set the parts | 3,409 | 3,434 | https://www.youtube.com/watch?v=McV4a6umbAY&t=3409s | AI for Imperfect-Information Games: Beating Top Humans in No-Limit Poker | |
McV4a6umbAY | that I don't necessarily want to model, because I'm very risk averse. So, that could become like adversarial moves, and I to be robust. Have you thought about how well other techniques would apply, or completely out of the ballpark. >> To be honest, I never really considered that direction for the research. I think there's a lot of potential here. This is an area of research that I think has | 3,434 | 3,458 | https://www.youtube.com/watch?v=McV4a6umbAY&t=3434s | AI for Imperfect-Information Games: Beating Top Humans in No-Limit Poker | |
McV4a6umbAY | been overlooked by a lot of people. So, it's actually been a very small community that has been working on the space, and I think people are starting to appreciate that can be applied to a lot of different things. We're starting to see papers on how to apply something like counterfactual regret minimization to more mainstream reinforcement learning topics. So, I think there was a paper from Berkeley recently on | 3,458 | 3,479 | https://www.youtube.com/watch?v=McV4a6umbAY&t=3458s | AI for Imperfect-Information Games: Beating Top Humans in No-Limit Poker | |
McV4a6umbAY | regret minimization in single-agent settings. So, I think there is definitely potential to extend the research and to the more traditional reinforcement learning settings, but I have not looked into that yet. >> [inaudible] a little bit, is there anything I'm learning? So, in the real world one of the things that's really difficult is, I don't know the rules of the game | 3,479 | 3,504 | https://www.youtube.com/watch?v=McV4a6umbAY&t=3479s | AI for Imperfect-Information Games: Beating Top Humans in No-Limit Poker | |
McV4a6umbAY | and I really have no idea what my problems hails are often. Is that something that people have looked at? Trying to basically think about simultaneously trying to improve my own hails trying to get them in model of what's going on with my opponent. >> Yes. I guess that would be kind of a subset of the previous case we discussed, where it will be like, maybe you know | 3,504 | 3,526 | https://www.youtube.com/watch?v=McV4a6umbAY&t=3504s | AI for Imperfect-Information Games: Beating Top Humans in No-Limit Poker | |
McV4a6umbAY | the structure of the game but you don't know what the opponent's payoffs. I think this has been looked at in the game theory community, but more in simple cases not like large-scale computer science settings. So, I could be wrong about that I don't know too well. >> As you said I should know the game theory, yes sure. So, there are models where there's a point in time but it's | 3,526 | 3,551 | https://www.youtube.com/watch?v=McV4a6umbAY&t=3526s | AI for Imperfect-Information Games: Beating Top Humans in No-Limit Poker | |
IeaD0ZaUJ3Y | you hi I'm professor Pete Carr I'm been at the University of Minnesota for about thirty-eight years I've worked with many many graduate students in class and in my research lab and I find it useful to work with students um to teach them how to read a paper on this first slide that I want to show you is an outline of of the way a typical scientific paper is organized and I think most beginning | 0 | 53 | https://www.youtube.com/watch?v=IeaD0ZaUJ3Y&t=0s | How to Read a Paper Efficiently (By Prof. Pete Carr) | |
IeaD0ZaUJ3Y | students instinctively start reading a paper in in order as the paper is printed for instance they read the title then they go on to the abstract then they read the introduction and so on and so forth working their way from the beginning to the end of the article don't do this this is not a good use of your time there's a better way to do things which is what I'm going to tell you about | 53 | 84 | https://www.youtube.com/watch?v=IeaD0ZaUJ3Y&t=53s | How to Read a Paper Efficiently (By Prof. Pete Carr) | |
IeaD0ZaUJ3Y | today let me jump in here at this point and tell you about a fairly simple algorithm if you will about how to get the most out of a paper with with the least effort and I think to do this you have to think of reading a paper as a two-phase process in the first phase you're surveying the paper the article to see if it's really worth investing a lot of time in and this is the way you | 84 | 113 | https://www.youtube.com/watch?v=IeaD0ZaUJ3Y&t=84s | How to Read a Paper Efficiently (By Prof. Pete Carr) | |
IeaD0ZaUJ3Y | you keep up with what's going on in in the literature you'll probably have some sort of service that provides you with with papers based upon keywords which are by and large taken from the abstract of of the paper off so the first step to keep more first thing to keep in mind is that you're allowed to stop this process at any point when you become disinterested in going further next you | 113 | 148 | https://www.youtube.com/watch?v=IeaD0ZaUJ3Y&t=113s | How to Read a Paper Efficiently (By Prof. Pete Carr) | |
IeaD0ZaUJ3Y | will look undoubtedly at the key words and and the title of the paper if these don't interest you at all you stop next thing to look at really is the abstract it's the most important part of the paper for getting acquainted with the paper but next I think you want to jump to the conclusions you don't read the intermediate steps you don't look at the experimental and the introduction and | 148 | 176 | https://www.youtube.com/watch?v=IeaD0ZaUJ3Y&t=148s | How to Read a Paper Efficiently (By Prof. Pete Carr) | |
IeaD0ZaUJ3Y | the results and discussion look at the conclusions because if the conclusions are not relevant to you probably you don't want to go any further so that this basically at this point you've surveyed the paper and you know whether or not it's really worth your while to invest any time on it the next thing that I think it's best to do again because it's fairly fast is to | 176 | 202 | https://www.youtube.com/watch?v=IeaD0ZaUJ3Y&t=176s | How to Read a Paper Efficiently (By Prof. Pete Carr) | |
IeaD0ZaUJ3Y | take a good look at the tables and the figures and the captions because you can do this quickly and it will tell you the main things that went on when the when the scientists did their work and again it will help you decide do I really want to dig into this paper or not if that's the case and you want to dig in then the place to start is the introduction and | 202 | 230 | https://www.youtube.com/watch?v=IeaD0ZaUJ3Y&t=202s | How to Read a Paper Efficiently (By Prof. Pete Carr) | |
IeaD0ZaUJ3Y | this unit have to start reading seriously and the introduction will provide you with essential background information that's one of the roles of the introduction another role of the introduction is to let you know why the authors of the paper did did the particular study and I think these are important things for you to know before digging in the real heart of a paper is | 230 | 259 | https://www.youtube.com/watch?v=IeaD0ZaUJ3Y&t=230s | How to Read a Paper Efficiently (By Prof. Pete Carr) | |
IeaD0ZaUJ3Y | the results and discussion section of the paper here's where you're going to spend most of your time and going through the paper finally at this point you may decide to stop you've had enough but if the paper is really extremely relevant to what you are currently working on then it's time to dig very deeply and get into the details of the experimental section of | 259 | 285 | https://www.youtube.com/watch?v=IeaD0ZaUJ3Y&t=259s | How to Read a Paper Efficiently (By Prof. Pete Carr) | |
IeaD0ZaUJ3Y | the paper and this is where you really learn what the authors actually did but more importantly it's how you it's where you learn how they did things and you may need that level of detail in your own work um once once you've done reading the paper uh you can you can stop however I think it's a really good idea to develop some kind of system where you take some notes on the paper | 285 | 319 | https://www.youtube.com/watch?v=IeaD0ZaUJ3Y&t=285s | How to Read a Paper Efficiently (By Prof. Pete Carr) | |
IeaD0ZaUJ3Y | these notes aren't going to serve you any good next week or maybe even next month but down the road maybe when you start to write your own first paper having some notes on these papers that you've read will be very beneficial and will really save you a lot of time it'll tell you which papers you should reread before you start writing which papers you don't need to include in as | 319 | 349 | https://www.youtube.com/watch?v=IeaD0ZaUJ3Y&t=319s | How to Read a Paper Efficiently (By Prof. Pete Carr) | |
IeaD0ZaUJ3Y | references in your own manuscript because they're not relevant so taking some notes when you can as you finish reading a paper is a really good idea and and these notes should be in a notebook or in some system and not simply written on the PDF of the paper because that you can't collect those those notes very easily you want these notes readily accessible for instance on | 349 | 377 | https://www.youtube.com/watch?v=IeaD0ZaUJ3Y&t=349s | How to Read a Paper Efficiently (By Prof. Pete Carr) | |
IeaD0ZaUJ3Y | an index card or a bunch of index cards so that you can flip through them quickly and take a look at all the all the relevant papers there's an old saying I think it's Chinese in origin that the faintest writing is better than the best memory and in the course of my time involved in science I am here to tell you that's that old saying is really right these are the main things I | 377 | 408 | https://www.youtube.com/watch?v=IeaD0ZaUJ3Y&t=377s | How to Read a Paper Efficiently (By Prof. Pete Carr) | |
IeaD0ZaUJ3Y | wanted to leave you with reading a paper or surveying a paper may take 30 seconds or if it's really relevant it may take several days several hours or even the major part of a day to go through the entire manuscript and really come to grips with it that's that's about all I can tell you it's a it's a skill that has to be learned to be productive in graduate school and productive as a | 408 | 437 | https://www.youtube.com/watch?v=IeaD0ZaUJ3Y&t=408s | How to Read a Paper Efficiently (By Prof. Pete Carr) | |
_p8vFSUesNs | thank you very much Michael very much for the invitation it's a great pleasure to be here and we're in some sense more of a user of many of the deep learning techniques which have been developed here in this community and I just wanted to highlight a few examples of how we can use deep learning and medical imaging and more specifically talk about image reconstruction super-resolution and | 0 | 34 | https://www.youtube.com/watch?v=_p8vFSUesNs&t=0s | Daniel Rueckert: "Deep learning in medical imaging" | |
_p8vFSUesNs | segmentation but probably in the same spirit as in as a last talk just as a sort of introduction there is a lot of excitement in in medical imaging in this in this area but there is also a lot of hype so if you look at at several magazine covers over the last last year actually there's you two amount of excitement for me is if you really read the headlines there's also quite a lot | 34 | 64 | https://www.youtube.com/watch?v=_p8vFSUesNs&t=34s | Daniel Rueckert: "Deep learning in medical imaging" | |
_p8vFSUesNs | of over-excitement and in some sense the press has picked up on this and quite often taken a number of things out of context so if you look at for example one of the the comments which geoff hinton made in 2017 said well you should rather stop training radiologists and if you read the whole interview he said many more things with many caveats but of course this one single headline is more | 64 | 92 | https://www.youtube.com/watch?v=_p8vFSUesNs&t=64s | Daniel Rueckert: "Deep learning in medical imaging" | |
_p8vFSUesNs | exciting than actually quoting really what he said in detail but a very famous radiologist said something more sensible probably is just a question where the AI will really replace radiologist probably the answer's no but actually a radiologist who do AI will replace a radiologist who don't and I think that's very it's very good a comment now the other thing is you probably all you're | 92 | 121 | https://www.youtube.com/watch?v=_p8vFSUesNs&t=92s | Daniel Rueckert: "Deep learning in medical imaging" | |
_p8vFSUesNs | familiar with conferences such as nips and you've seen it and many many more people going to these conferences so some of our radiology colleagues go to a conference called RNA and if you think nips is big and very unmanageable our SNA has around 45 to 50,000 attendees in Chicago is apparently the only city in the world big enough to host this conference but one of the things which really | 121 | 149 | https://www.youtube.com/watch?v=_p8vFSUesNs&t=121s | Daniel Rueckert: "Deep learning in medical imaging" | |
_p8vFSUesNs | happened this year or last year 2017 which is very interesting is normally if you go to these conferences you have a strand on x-ray stand on CT on molecular imaging it's basically all the different imaging modality and this year for the first time an equal strand is machine learning which really shows you how much machine learning is changing medicine in particular probably medical imaging | 149 | 178 | https://www.youtube.com/watch?v=_p8vFSUesNs&t=149s | Daniel Rueckert: "Deep learning in medical imaging" | |
_p8vFSUesNs | there are huge number of opportunities many of which we sort of copied effectively from the vision area but another another really important aspect is there's a huge amount of data which is now publicly available in medical imaging and I just want to highlight one particular example because I've been involved in in this in the UK is something called the UK biobank for | 178 | 205 | https://www.youtube.com/watch?v=_p8vFSUesNs&t=178s | Daniel Rueckert: "Deep learning in medical imaging" | |
_p8vFSUesNs | those of you who don't know what UK biobank is actually it's a population study of nearly 500,000 people but a hundred thousand of these subjects will be imaged and actually they have already done 1/5 so they've already done over 20,000 of these subjects and so they're acquiring very high-resolution imaging data from from this subject and one of the things which is interesting is | 205 | 231 | https://www.youtube.com/watch?v=_p8vFSUesNs&t=205s | Daniel Rueckert: "Deep learning in medical imaging" | |
_p8vFSUesNs | they're not only acquiring images of the whole body which of course is very useful with MRI but they're also acquiring a dedicated brain imaging so not only structural brain imaging functional brain imaging and diffusion brain imaging and they're also quite dynamic images of the heart for example are allowing you to look at cardiovascular function and for example | 231 | 255 | https://www.youtube.com/watch?v=_p8vFSUesNs&t=231s | Daniel Rueckert: "Deep learning in medical imaging" | |
_p8vFSUesNs | study the interaction between the brain and the heart which is something which is not very well understood at the moment and more importantly you also have available lifestyle information how many cross wants to eat how many cups of coffee you drink genetics and Clint and links to clinical records now why is this exciting is because actually that data set is available | 255 | 279 | https://www.youtube.com/watch?v=_p8vFSUesNs&t=255s | Daniel Rueckert: "Deep learning in medical imaging" | |
_p8vFSUesNs | everybody in the world to use so if you wanted to use that data set you have to fill out an application make sure that you don't try to be anonymized data but otherwise you can download the data and and use it for for research purposes so that's it for example really something which is a game changer I think in medicine so if you look at what are the opportunities more specifically really | 279 | 306 | https://www.youtube.com/watch?v=_p8vFSUesNs&t=279s | Daniel Rueckert: "Deep learning in medical imaging" | |
_p8vFSUesNs | sort of there is a sort of pyramid where the value of what we do increases the sort of lowest level you can use machine learning to help you with image reconstruction for example you can automatically plan your scans so if you go if you line in mo a scanner for example there's quite a lot of fiddling by an operator by not lo radiologist but typically a radiographer | 306 | 329 | https://www.youtube.com/watch?v=_p8vFSUesNs&t=306s | Daniel Rueckert: "Deep learning in medical imaging" | |
_p8vFSUesNs | who sort of plans the imaging you can of course use image enhancement super resolution is something which I'll show you in a moment you can do the conventional semantic image interpretation for example find organs Sigma and these organs you can quantify biomarkers for example measure a tumor volume and then sort of if you come to the more higher level of the more we're | 329 | 355 | https://www.youtube.com/watch?v=_p8vFSUesNs&t=329s | Daniel Rueckert: "Deep learning in medical imaging" | |
_p8vFSUesNs | actually probably highest added value is sort of computer-aided interpretation and diagnosis there are probably very few applications at the moment really being used and probably the only one I can really think of where actually machine learning has had an impact at those high level features is at the moment in for example mammography screenings so there are some systems | 355 | 379 | https://www.youtube.com/watch?v=_p8vFSUesNs&t=355s | Daniel Rueckert: "Deep learning in medical imaging" | |
_p8vFSUesNs | which effectively act as a second reader to a radiologist but really we haven't really cracked the top of this pyramid but there's quite a lot of work going on at the bottom I also want to show you a few challenges where I think really we we need your help in developing new methods sort of at the moment most of the techniques we use are supervised techniques so that means our training | 379 | 409 | https://www.youtube.com/watch?v=_p8vFSUesNs&t=379s | Daniel Rueckert: "Deep learning in medical imaging" | |
_p8vFSUesNs | data is absolutely critical for what we do if you look at our colleagues envision what they typically do is well you go crowdsource your your labels or your annotations and if you say that to radiologists they don't really react very well to this so they they obviously saying well she'd really asked some experts to help you with we're doing that but in the UK at least for example | 409 | 435 | https://www.youtube.com/watch?v=_p8vFSUesNs&t=409s | Daniel Rueckert: "Deep learning in medical imaging" | |
_p8vFSUesNs | we already trained far too few radiologists so if you ask them to generate training data ie label data it's really very challenging and quite difficult to do and more importantly and this is something which is quite I think not always perfectly understood is if I ask an observer to tell me whether there's a car in a photo or not then the answer is pretty unambiguous either | 435 | 461 | https://www.youtube.com/watch?v=_p8vFSUesNs&t=435s | Daniel Rueckert: "Deep learning in medical imaging" | |
_p8vFSUesNs | there is a car or there isn't in in medical imaging our training data is often imperfect ie if you ask three different radiologists that will give you hopefully not three different answers but they might give you more than one answer so you really need to train in a scenario where your data might not be perfectly labeled and of course if that's then not perfectly | 461 | 484 | https://www.youtube.com/watch?v=_p8vFSUesNs&t=461s | Daniel Rueckert: "Deep learning in medical imaging" | |
_p8vFSUesNs | labeled how do you actually validate your algorithm against something where there is really no gold standard the other thing is quite often is we've had very great success are reporting fantastic algorithms in a scientific paper and then when you deploy them in a clinical scenario they don't really work that well and if you actually look at scanners they're really only three big | 484 | 511 | https://www.youtube.com/watch?v=_p8vFSUesNs&t=484s | Daniel Rueckert: "Deep learning in medical imaging" | |
_p8vFSUesNs | scanner manufacturers and when they produce scanners they're not only very in the color as you see here but they really produce slightly different images to a human observer that doesn't really matter but actually turns out to most of our models we train with machine learning it actually matters and quite often we don't have access to data from all different sites or from all | 511 | 535 | https://www.youtube.com/watch?v=_p8vFSUesNs&t=511s | Daniel Rueckert: "Deep learning in medical imaging" | |
_p8vFSUesNs | different vendors and if you every time you then go to your radiologist colleague and tell them well just annotate a few more data set and then I can do some retraining they don't really react very well to this at least in my experience okay so I want to show you three different applications for where we at the moment use deep learning and we're actually has | 535 | 560 | https://www.youtube.com/watch?v=_p8vFSUesNs&t=535s | Daniel Rueckert: "Deep learning in medical imaging" | |
_p8vFSUesNs | had significant impact really transformed quite a lot of the things we do and the first one is really on image reconstruction and and this is really probably something which a number of you will have worked on it's actually quite well understood problem using it as an inverse problem but I want to show you a particular application here and that is using a modality which some of you may | 560 | 587 | https://www.youtube.com/watch?v=_p8vFSUesNs&t=560s | Daniel Rueckert: "Deep learning in medical imaging" | |
_p8vFSUesNs | know is called magnetic resonance imaging it's a great modality because it's it's safe it can do a lot of things can show you many different properties of the body but it's relatively slow that means it's good for measuring or acquiring images of the brain but not so good if you want to for example do cardiac imaging which you see here so if you look at these cardiac images which | 587 | 612 | https://www.youtube.com/watch?v=_p8vFSUesNs&t=587s | Daniel Rueckert: "Deep learning in medical imaging" | |
_p8vFSUesNs | you see on this particular slide this looks like it's one single heartbeat but unfortunately mr imaging is not fast enough to acquire this so what you typically do is you measure with an ECG in which actor in which state the heart is and then I'm actually taking bits of data from different heartbeats and I'm assuming that if my ECG signal shows me I'm in the same hearts state as as | 612 | 640 | https://www.youtube.com/watch?v=_p8vFSUesNs&t=612s | Daniel Rueckert: "Deep learning in medical imaging" | |
_p8vFSUesNs | previously then I can take these measurements and combine them to form an image so this particular image here is typically probably an average of half I had 10 different heart beats so it's not a single heartbeat be much nicer if we can acquire this faster because actually to acquire the state of the patient has to hold their breath and that's especially if you have heart disease | 640 | 663 | https://www.youtube.com/watch?v=_p8vFSUesNs&t=640s | Daniel Rueckert: "Deep learning in medical imaging" | |
_p8vFSUesNs | then of course it gets a bit more tricky so 10 heart beats is right around 8 to 10 seconds that's how long you have to hold your breath and you really would like to do this fast enough you don't have to hold your breath at all okay so just a very primitive explanation of how we take our measurements in MRI imaging and really the physics is not so important but it's just want to | 663 | 691 | https://www.youtube.com/watch?v=_p8vFSUesNs&t=663s | Daniel Rueckert: "Deep learning in medical imaging" | |
_p8vFSUesNs | highlight why this is so slow what we typically do is we traverse our measurement space in which we take measurements which is called the K space and we have to traverse it sequentially for a number of reasons which I don't really want to go into too much detail but once I've made my measurements in k space then actually image reconstruction is trivial I just applied the Fourier | 691 | 713 | https://www.youtube.com/watch?v=_p8vFSUesNs&t=691s | Daniel Rueckert: "Deep learning in medical imaging" | |
_p8vFSUesNs | transform and I get back my image so that is a relatively slow process because I have to acquire these measurements in k space sequentially and of course if I want to for example create dynamic images of the heart I have to keep on doing the same thing over and over again but actually the heart is only changing a bit between these different acquisitions so there's | 713 | 738 | https://www.youtube.com/watch?v=_p8vFSUesNs&t=713s | Daniel Rueckert: "Deep learning in medical imaging" | |
_p8vFSUesNs | an enormous amount of spatial temporal redundancy in the data now what is the easiest solution to accelerate your imaging process well instead of acquiring all of the measurements I need i just acquire a subset of the measurements which i need so for example if i acquire as you see here in the bottom only 25% of the data of course i'm twenty and four times faster which | 738 | 767 | https://www.youtube.com/watch?v=_p8vFSUesNs&t=738s | Daniel Rueckert: "Deep learning in medical imaging" | |
_p8vFSUesNs | is great for the patient but for the radiologist this image is not really acceptable because it looks actually much degraded because i have a lot of aliasing artifacts in the data so this problem has been well studied there are many techniques which can try to help you recover that that information which is in the top image you want to effectively denoise or D alias this | 767 | 793 | https://www.youtube.com/watch?v=_p8vFSUesNs&t=767s | Daniel Rueckert: "Deep learning in medical imaging" | |
_p8vFSUesNs | image here at the bottom and the most commonly used techniques are effectively compressed sensing techniques so these compressed sensing techniques have been around for a while and MRI imaging they've been around for ten years very successfully used but more recently some some machine learning techniques have really significantly outperformed these techniques and really what is the | 793 | 817 | https://www.youtube.com/watch?v=_p8vFSUesNs&t=793s | Daniel Rueckert: "Deep learning in medical imaging" | |
_p8vFSUesNs | key difference between the compressed sensing techniques you see in the in the top and those here at the bottom so the ones in the top they effectively use generic priors sparsity low ranked they're not really data-driven priors whereas the techniques in the bottom they effectively try to learn the price from the data and try to improve our acquisition in that bad sense so just | 817 | 844 | https://www.youtube.com/watch?v=_p8vFSUesNs&t=817s | Daniel Rueckert: "Deep learning in medical imaging" | |
_p8vFSUesNs | want to quickly show you the problem formulation we have we have effectively our case based measurements so all our measurements are because we're making frequency measurements or complex values so the images we typically see are just the magnitude of those complex numbers but our measurements are complex valued and the image we want to recover is also complex valued and in fact we our | 844 | 870 | https://www.youtube.com/watch?v=_p8vFSUesNs&t=844s | Daniel Rueckert: "Deep learning in medical imaging" | |
_p8vFSUesNs | measurements are related by to the image through a undersampled for you including matrix so this effectively under samples your measurement space and applies a Fourier transform and of course our acquisition noise and the under samples for your operation if you want to write it down differently you can just write it down as a Fourier operation and then effectively a mask which defines how | 870 | 898 | https://www.youtube.com/watch?v=_p8vFSUesNs&t=870s | Daniel Rueckert: "Deep learning in medical imaging" | |
_p8vFSUesNs | you're under sampled case space and that's that's quite important okay so in what you're trying to then solve mathematically is typically an unconstrained optimization problem consisting of your regularization term if you would use compressed sensing you'd probably use here an l1 or LZ row prior a norm or and the data fidelity term and that data fidelity term is | 898 | 924 | https://www.youtube.com/watch?v=_p8vFSUesNs&t=898s | Daniel Rueckert: "Deep learning in medical imaging" | |
_p8vFSUesNs | quite crucial and it's actually something which is not so easy to bring into into the equation if you use a machine learning so you know if we try to for example use a CNN because in some sense we we're trying to formulate this as a Dean problems with taking as an input one image and we're trying to produce a denounced image then we effectively have two different terms first which we're | 924 | 953 | https://www.youtube.com/watch?v=_p8vFSUesNs&t=924s | Daniel Rueckert: "Deep learning in medical imaging" |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.