WEBVTT 1 00:00:05.220 --> 00:00:06.839 WH11NE- Sunrise: Thank you, everybody for coming to. 2 00:00:19.199 --> 00:00:20.490 WH11NE- Sunrise: This analysis. 3 00:00:21.780 --> 00:00:25.050 WH11NE- Sunrise: To see a fellow and. 4 00:00:36.450 --> 00:00:39.600 WH11NE- Sunrise: Everyone, thank you for the invitation for the farm. 5 00:00:41.040 --> 00:00:41.460 WH11NE- Sunrise: Yes. 6 00:00:43.650 --> 00:00:46.530 WH11NE- Sunrise: She learning for the simulations. 7 00:00:55.320 --> 00:01:05.190 WH11NE- Sunrise: Yes, all i'll go over briefly ratio thc and wearing machine learning could fit in the motivation for it i'll then discuss. 8 00:01:05.700 --> 00:01:17.370 WH11NE- Sunrise: The generative models and metrics evaluate them as a representative data to us with them and then just give an overview of current work and our approach call and he can. 9 00:01:18.600 --> 00:01:20.310 Some some discussion. 10 00:01:21.450 --> 00:01:25.830 WH11NE- Sunrise: And so yeah it's quickly to go over the words for using machine learning. 11 00:01:32.580 --> 00:01:34.290 WH11NE- Sunrise: That was interesting. 12 00:01:35.850 --> 00:01:49.980 WH11NE- Sunrise: yeah so detector simulation takes roughly 40% of cpu resources at 70%. 13 00:01:54.390 --> 00:01:54.900 WH11NE- Sunrise: So. 14 00:02:00.270 --> 00:02:00.690 WH11NE- Sunrise: So it's. 15 00:02:02.070 --> 00:02:02.580 WH11NE- Sunrise: outdated. 16 00:02:03.630 --> 00:02:04.350 WH11NE- Sunrise: As a. 17 00:02:05.910 --> 00:02:25.170 WH11NE- Sunrise: Calm we're going to need at least an order of magnitude more simulations to keep up the statistics and we're improving or actors to the more granular increased complexity and recruiting we're going to face some technical challenges and perhaps machine learning is a possible solution. 18 00:02:26.910 --> 00:02:33.150 WH11NE- Sunrise: And, in turn, the word can fit in i'll just kind of go over the the pipeline. 19 00:02:34.590 --> 00:02:39.810 WH11NE- Sunrise: You know bias towards the investor simulation, but I think the actually clickable. 20 00:02:40.950 --> 00:02:46.020 WH11NE- Sunrise: elements which are usually start off with some kind of hard process. 21 00:02:46.950 --> 00:02:56.520 WH11NE- Sunrise: you're trying to the new age and using the showering and ization of delaying event action and you end up with a list of can read your articles. 22 00:02:57.060 --> 00:03:08.340 WH11NE- Sunrise: That you then traverse through the detector thing machines interactions to have little bits new model, the digitization and construction you get reconstructing it. 23 00:03:08.850 --> 00:03:20.310 WH11NE- Sunrise: By run it through, for example, particle flow to get articles and then find the bedtime feature instant analyses and turn the machine learning, there are a lot of opportunity where we can. 24 00:03:21.900 --> 00:03:32.340 WH11NE- Sunrise: incorporate it so, for example, I just do to do the gen level stuff, it has to be replaced GM for them, or if. 25 00:03:33.600 --> 00:03:52.320 WH11NE- Sunrise: You could just do a parametric responsive detector typical model, the response to individual articles and go really wild and just everything end to end and you basically have a bunch of trade offs you're making so you're trading the accuracy we're doing a full simulation Jan. 26 00:03:53.640 --> 00:03:59.550 WH11NE- Sunrise: To get an improvement in speed and you're also trading off their. 27 00:04:00.840 --> 00:04:02.520 WH11NE- Sunrise: interests and having your process. 28 00:04:04.410 --> 00:04:07.980 WH11NE- Sunrise: really try something which just go directly to high level. 29 00:04:09.030 --> 00:04:09.420 In. 30 00:04:10.650 --> 00:04:11.310 WH11NE- Sunrise: This article. 31 00:04:14.520 --> 00:04:19.950 WH11NE- Sunrise: This is something we have to think about it, you know, try to meet you in the top of ways to trust. 32 00:04:21.480 --> 00:04:32.700 WH11NE- Sunrise: I rely on these galaxies but i'll just give the definition for this task machine learning language, where you're trying to model this data. 33 00:04:33.360 --> 00:04:43.440 WH11NE- Sunrise: You have with the model T data for the underlying distribution your exes different any of these steps in the in the pipeline. 34 00:04:44.250 --> 00:05:00.540 WH11NE- Sunrise: And this is a really rich area and machine learning it's called peak generated modeling where the deep comes from using deep neural networks, which are very flexible and expressive and typically try to penetrate this model effects, so it. 35 00:05:02.520 --> 00:05:11.550 WH11NE- Sunrise: Is all go more into these the candidate models stuff for me to pause here for questions or do you usually hits that question. 36 00:05:15.810 --> 00:05:17.460 WH11NE- Sunrise: Okay, you can just interrupt me anytime. 37 00:05:19.800 --> 00:05:22.950 WH11NE- Sunrise: Okay, so no question so i'll go over some. 38 00:05:24.690 --> 00:05:31.680 WH11NE- Sunrise: types of the generative model so it's a it's a you know, a different areas missionary than your typical use case where you. 39 00:05:32.430 --> 00:05:52.410 WH11NE- Sunrise: Have a fix internally there and clear out for you're trying to create a generative modeling the little more difficult to get created the first class of models are talking about for doing this or direct models were basically you're trying to explicitly find the distribution. 40 00:05:54.210 --> 00:06:07.440 WH11NE- Sunrise: network and the first model sinner auto regressive models, where the idea is you make a very strong model assumption, which is a each component of your data. 41 00:06:09.030 --> 00:06:14.010 WH11NE- Sunrise: individually, a sample can be predicted entirely by the previous data points. 42 00:06:16.740 --> 00:06:19.320 WH11NE- Sunrise: By entirely by the previous data points. 43 00:06:20.700 --> 00:06:25.170 WH11NE- Sunrise: And can enter the output, the next component of. 44 00:06:26.220 --> 00:06:41.460 WH11NE- Sunrise: Your data sample by learning a distribution learning to predict the next level this at all the previous all the previous episodes, this is a very strong assumption, which may not necessarily apply in our case it's. 45 00:06:42.690 --> 00:06:59.220 WH11NE- Sunrise: Not necessarily the best model but it's a nice way to get the explicitly included another popular type of model is what are what are called flow based models so it's really just that changing variable formula, where you try to find an easier. 46 00:07:00.360 --> 00:07:06.450 WH11NE- Sunrise: way to present tense or transform next into a simpler distribution the backwards, a. 47 00:07:08.940 --> 00:07:15.090 WH11NE- Sunrise: The backward parts, we have to train the model and you try to go from the simpler or go from this. 48 00:07:16.200 --> 00:07:35.340 WH11NE- Sunrise: potentially very complicated distribution on your data effects and try to transform it deter that lead to a simpler description on Z and then the key is that new every transformation is or isn't your sample from the simpler, distribution and came back to your. 49 00:07:37.110 --> 00:07:55.740 WH11NE- Sunrise: it's a very nice idea very simple conceptually another class model will talk about our fusion based models, which are a little similar where you're trying to go from your real sample and then keep adding noise to get up to a gaussian distribution. 50 00:07:57.150 --> 00:08:08.730 WH11NE- Sunrise: And then for for actually generating samples you learn the inverse process you try to learn it and network that we're trying to do is just real simple. 51 00:08:10.380 --> 00:08:13.410 WH11NE- Sunrise: And so, these are the types of models, where you get. 52 00:08:14.430 --> 00:08:29.400 WH11NE- Sunrise: try to get the effects are very nice to have lots of benefits like you have access to the exact like today, we have a very simple training, so you just try to minimize and make it as long as we read very stable, so all these nice properties. 53 00:08:30.570 --> 00:08:37.500 WH11NE- Sunrise: But unfortunately in practice they're actually typically have performed by models called can which we'll talk about which channel. 54 00:08:41.190 --> 00:09:03.360 WH11NE- Sunrise: So again, our titles Beijing very good variable our the ideas use humor high dimensional X can be characterized entirely by lower dimensional in features be and the general process is simple you sample from Z would presumably had simpler prior and the top and learn. 55 00:09:06.690 --> 00:09:17.670 WH11NE- Sunrise: And first type of variation will auto with quarters, where you still are trying to maximize the likelihood, but actually an approximation to a lot like. 56 00:09:18.660 --> 00:09:24.690 WH11NE- Sunrise: All the evidence based more balanced, which has these two terms you're missing a kale that virgins. 57 00:09:25.650 --> 00:09:37.500 WH11NE- Sunrise: In this equation, but these two terms can be implemented nicely in this particular auto encoder so you can get this kale divergence by kind to encode probabilistic we. 58 00:09:37.950 --> 00:09:46.980 WH11NE- Sunrise: sampled into your space, but some assumption upon prior So here we are, and you just measure the kale divergence. 59 00:09:47.790 --> 00:10:00.690 WH11NE- Sunrise: Between your encoded distribution and the prior and then you try to reconstruct the sample back to tighter get the real simple logistically and that gives you your construction lots in word. 60 00:10:04.530 --> 00:10:08.970 WH11NE- Sunrise: And then the next model which are called the genitive adversary on that works. 61 00:10:10.260 --> 00:10:12.180 WH11NE- Sunrise: Is kind of bizarrely just. 62 00:10:13.350 --> 00:10:15.660 WH11NE- Sunrise: As likely good approach entirely. 63 00:10:16.920 --> 00:10:30.510 WH11NE- Sunrise: Instead, for your boss you actually use another word all to discriminate or to try and distinguish between real life examples, the disclaimer and either the training the generators your training it to learn whether it's. 64 00:10:33.720 --> 00:10:40.650 WH11NE- Sunrise: Basically, creating a feedback loop with each other, each network, improving each other effectively an adversary really. 65 00:10:41.940 --> 00:10:52.920 WH11NE- Sunrise: And what this means is that it's pretty hard to train with you have to balance the networks together each step and you lose access to the likelihood adversarial trainings but it's. 66 00:10:54.300 --> 00:10:56.640 WH11NE- Sunrise: difficult, but when you. 67 00:10:57.870 --> 00:10:59.010 WH11NE- Sunrise: With enough patients. 68 00:11:00.300 --> 00:11:07.230 WH11NE- Sunrise: were actually able to get better results, and all these other models at least that's what we've seen in computer vision fields. 69 00:11:07.740 --> 00:11:17.160 WH11NE- Sunrise: Until very recently until just last year, where a different type of model of score based infusion models actually started eating gans image generator. 70 00:11:17.640 --> 00:11:36.120 WH11NE- Sunrise: So just December last year that's very exciting it's basically modeling the gradient of the problem, a lot of like weird if that is the comfortability distribution itself that are ahead of time to go into the chocolate it's a very interesting area to pursue. 71 00:11:37.800 --> 00:11:40.260 WH11NE- Sunrise: Instead i'll go over to move on to. 72 00:11:41.370 --> 00:12:01.890 WH11NE- Sunrise: Probably the big questions you have when it comes to generate modeling buying into physics, which is how do we actually trust the data which comes out of the model that we use and in our analyses and related leads, how do we compare the different, how do we choose one model for our. 73 00:12:03.060 --> 00:12:18.600 WH11NE- Sunrise: simulation into the answer to both of the this the US well on our developed evaluation metrics and go over some criteria we want for these metrics and then propose some for us cms. 74 00:12:19.830 --> 00:12:36.570 WH11NE- Sunrise: So we want metrics which are sensitive to the quality of your data, the diversity of the data is learning the full distribution or gets samples, and ultimately physics performance, so can we use this system can reinterpret the mentioned w. 75 00:12:37.770 --> 00:12:38.310 WH11NE- Sunrise: In a way, which. 76 00:12:39.750 --> 00:12:47.220 WH11NE- Sunrise: And one, the more practical side to actually compare between all these different models are coming out, we also use metrics to be standardized. 77 00:12:48.510 --> 00:12:58.950 WH11NE- Sunrise: reproducible that we measure this metric the next day will be the same value and efficient should be fast enough to be able to validate all. 78 00:13:01.710 --> 00:13:02.220 WH11NE- Sunrise: So the. 79 00:13:03.630 --> 00:13:12.000 WH11NE- Sunrise: These metrics and we're coming up with metrics and probably most important aspect of generative model otherwise it's hard to actually use them. 80 00:13:13.230 --> 00:13:18.870 WH11NE- Sunrise: And so we propose to key metrics which I think in conjunction said it's by all these properties. 81 00:13:20.250 --> 00:13:20.700 WH11NE- Sunrise: One is. 82 00:13:21.960 --> 00:13:29.040 WH11NE- Sunrise: One washer sign distance between the two distribution, so this is coming from more specific set of things than one is anything. 83 00:13:30.750 --> 00:13:42.930 WH11NE- Sunrise: You want to establish a great already in machine learning, which is called the appreciate inception distance it's just the standard metric being used in computer vision. 84 00:13:44.490 --> 00:13:47.460 WH11NE- Sunrise: days and we're trying to adapt it to. 85 00:13:50.010 --> 00:13:50.610 WH11NE- Sunrise: and 86 00:13:51.630 --> 00:14:06.810 WH11NE- Sunrise: Like I said practically we found it to be very excited so it's all six of those criteria and in an effort to standardize these metrics goals to release them, you know Open Source package for everybody to use models. 87 00:14:08.460 --> 00:14:31.050 WH11NE- Sunrise: So i'll go over the out the first is the the one washer same distance, so I guess it's coming from the traditional physics approach, which is to compare physical distributions this is used for everything so not just machine learning for fellow evaluation fast and. 88 00:14:32.490 --> 00:14:35.250 WH11NE- Sunrise: machine learning we get images. 89 00:14:36.390 --> 00:14:39.090 WH11NE- Sunrise: So our proposal is to just decided this. 90 00:14:41.940 --> 00:14:48.360 WH11NE- Sunrise: Distance which is not work, you have to do to move one distribution to the other one. 91 00:14:49.860 --> 00:14:58.080 WH11NE- Sunrise: and have a lot of benefits to evaluate a lot of different physical features simultaneously, and this needs to. 92 00:14:59.280 --> 00:15:03.150 WH11NE- Sunrise: Greece a metric very sensitive to the quality of your samples. 93 00:15:04.290 --> 00:15:14.220 WH11NE- Sunrise: it's also very sensitive to model but different supports both mode collapse if a model just producing this sample and nostrils efficient. 94 00:15:15.540 --> 00:15:19.380 WH11NE- Sunrise: and fast it's reproducible its interpreted bull and also. 95 00:15:20.400 --> 00:15:32.430 WH11NE- Sunrise: You can use bootstrapping to measure the scores between just samples of just between just feel samples alone, if that gives you a baseline score to compare your models to which is also very useful. 96 00:15:34.470 --> 00:15:46.800 WH11NE- Sunrise: you're there are some console with this, which is that its beginning and then it also means that if you try to measure correlations between the features, you get into the curse of dimensionality scales. 97 00:15:47.460 --> 00:15:54.030 WH11NE- Sunrise: polynomial with the number of dimensions want opinion so it's not it's not scalable at all and. 98 00:15:55.470 --> 00:16:00.120 WH11NE- Sunrise: it's also not clear how we would activate the scores for individual teachers together. 99 00:16:01.200 --> 00:16:08.040 WH11NE- Sunrise: And it's all of these issues are actually solved by machine learning version of this yeah try to ask them before you go on so. 100 00:16:08.940 --> 00:16:25.080 WH11NE- Sunrise: you're talking about doing proceeding this distance on histograms yeah it's like it's not the underlying current goals or whatever that are used to it's not some abstract space for that are they talking about like taking a bit of making it smaller and making other. 101 00:16:26.760 --> 00:16:38.070 WH11NE- Sunrise: Higher in this transformation process, oh no the model themselves could be predicting that the parcels the model could be predicting the articles themselves either like a future with derived. 102 00:16:38.640 --> 00:16:56.850 WH11NE- Sunrise: From the prediction So even if you're predicting individual articles you, what is the distance what what space does this distance act on the distance is between the histogram sorry for the interruption of the feature get from the gents you produce it from the shoulders. 103 00:17:00.810 --> 00:17:08.130 WH11NE- Sunrise: But it's like instagram of what like above some high level feature that you that you decide is valuable or. 104 00:17:09.270 --> 00:17:25.650 WH11NE- Sunrise: instagram of the particles psp why PV they follow them it's you wouldn't want to look at it as much as possible the in the slide and yet for the jets on show parties, as well as well as Jeff level features. 105 00:17:28.200 --> 00:17:34.440 WH11NE- Sunrise: This give you like a different answer practice like if i'm if i'm sorry, Mr graham's I know how to do that first. 106 00:17:36.210 --> 00:17:47.880 WH11NE- Sunrise: Are there things the same whether rms is the same, but there's use this and other critters like ready to get used to all the moments to give you a different answer, or does it give us information by. 107 00:17:49.380 --> 00:17:49.740 WH11NE- Sunrise: You mean you. 108 00:17:50.760 --> 00:17:57.120 WH11NE- Sunrise: walk into sense, instead of some of these other like classical hold events for probability distributions. 109 00:17:59.250 --> 00:18:03.600 WH11NE- Sunrise: And then you can use as many as you want to. 110 00:18:05.340 --> 00:18:12.000 WH11NE- Sunrise: I don't know if this is better than looking at some me it doesn't give you more information than just the meat. 111 00:18:13.260 --> 00:18:17.520 WH11NE- Sunrise: But Yet if you use every all the moments, and maybe together. 112 00:18:19.260 --> 00:18:24.780 WH11NE- Sunrise: it'll be as usable by the show that it doesn't have the Nice property over some other other. 113 00:18:26.070 --> 00:18:28.290 WH11NE- Sunrise: distributions are for one day, Sir quantities. 114 00:18:29.610 --> 00:18:39.690 WH11NE- Sunrise: yeah I guess so you're just comparing trick is wrapped candy was just like three in Baghdad it just takes a big difference, like the iceberg yeah yeah oh. 115 00:18:40.710 --> 00:18:56.640 WH11NE- Sunrise: yeah there's this slide which other shoe it's a reasonable way to do it, a lot of people who use just crank squared or some of the goodness of fit the thing I think the benefit of using one, there is no resistance that that it takes the metrics these into account. 116 00:18:58.080 --> 00:19:02.250 WH11NE- Sunrise: Okay, maybe i'll show the example is besides get to that. 117 00:19:06.900 --> 00:19:07.710 WH11NE- Sunrise: Last point. 118 00:19:10.170 --> 00:19:14.700 WH11NE- Sunrise: There was a way to compute this distance without the bedding and and and. 119 00:19:15.720 --> 00:19:16.380 WH11NE- Sunrise: objects. 120 00:19:17.520 --> 00:19:18.030 WH11NE- Sunrise: That. 121 00:19:21.240 --> 00:19:23.550 WH11NE- Sunrise: i'm not sure like with a criminal. 122 00:19:24.690 --> 00:19:31.290 WH11NE- Sunrise: Well, because i'm like enough like washing gap right like you're comparing individual. 123 00:19:32.430 --> 00:19:32.670 WH11NE- Sunrise: Like. 124 00:19:33.780 --> 00:19:37.770 WH11NE- Sunrise: that's like a point that would be like equivalent to like a. 125 00:19:40.470 --> 00:19:51.450 WH11NE- Sunrise: threat right yeah it's not actually calculating the it's like it's using a lot that approximates right, but the thing looks like the. 126 00:19:52.770 --> 00:19:57.510 WH11NE- Sunrise: Like the restricting to like weights that are clipped right. 127 00:19:58.530 --> 00:20:13.380 WH11NE- Sunrise: That that kind of bonus it's I think it's using a network predict network to basically derive yeah the approximation of the this yet right, and in that case you don't need it. 128 00:20:15.570 --> 00:20:26.190 WH11NE- Sunrise: yeah your I guess your training and network to to take both sample samples from both distribution outlet the distance using a network yeah. 129 00:20:26.910 --> 00:20:38.580 WH11NE- Sunrise: yeah you could do that, and I think probably another that go another metric is to just view the network directly to just decide if they're the same distribution just kind of what you're doing. 130 00:20:39.840 --> 00:20:45.060 WH11NE- Sunrise: yeah that's another metric, which means that the caveat of that one is you only get. 131 00:20:46.140 --> 00:20:53.910 WH11NE- Sunrise: Some constant multiple at the distance the concept, so you have to call back at something maybe that's very sensitive this isn't something. 132 00:20:55.410 --> 00:21:01.320 WH11NE- Sunrise: You don't know when you're like you know when you're at the optimum within your training, but that doesn't give you like. 133 00:21:03.120 --> 00:21:06.930 WH11NE- Sunrise: That yeah yeah something you have to train a network. 134 00:21:10.200 --> 00:21:14.910 WH11NE- Sunrise: Maybe the bootstrapping can give you the calibration yeah but I was a little bit different. 135 00:21:18.300 --> 00:21:23.850 WH11NE- Sunrise: On the data you define purchasing damage that was the same distances just. 136 00:21:25.260 --> 00:21:28.650 WH11NE- Sunrise: what's the least amount of energy to move all the points in one day that's. 137 00:21:30.150 --> 00:21:33.090 WH11NE- Sunrise: All the points of the other data set but that. 138 00:21:34.350 --> 00:21:34.860 WH11NE- Sunrise: But the fact. 139 00:21:37.290 --> 00:21:44.340 WH11NE- Sunrise: But it scales fast, but the number of points you have been so you have smaller numbers it's so you have. 140 00:21:45.750 --> 00:21:51.420 WH11NE- Sunrise: Just as equal number of points at the midpoint of each of you so it's faster than the. 141 00:21:53.100 --> 00:21:54.240 WH11NE- Sunrise: cowboys distance. 142 00:21:56.550 --> 00:21:59.130 WH11NE- Sunrise: So it doesn't have to give any independent that that makes a lot easier. 143 00:22:00.540 --> 00:22:00.990 WH11NE- Sunrise: Okay. 144 00:22:02.460 --> 00:22:12.270 WH11NE- Sunrise: i'm sorry this is something that others have said, do, but just to be clear, the proposal is not complete by seeing you use w and distance. 145 00:22:13.980 --> 00:22:22.380 WH11NE- Sunrise: Distance on these tools and features, so that is what comprises the evaluation that's the yeah. 146 00:22:23.940 --> 00:22:27.030 WH11NE- Sunrise: You derive the features which you think are most important. 147 00:22:28.320 --> 00:22:28.740 WH11NE- Sunrise: candidate. 148 00:22:34.710 --> 00:22:35.070 WH11NE- Sunrise: Okay. 149 00:22:36.300 --> 00:22:48.480 WH11NE- Sunrise: So move on to the next, the other type, so this is like the machine learning version of this essentially the established or seems to be sackler Center. 150 00:22:49.620 --> 00:22:57.120 WH11NE- Sunrise: For computer vision which is basically instead of these features you directed by end or. 151 00:22:58.650 --> 00:23:09.960 WH11NE- Sunrise: acknowledge you use classifier instead so you train a classifier on your data set and you take the feature that learns. 152 00:23:11.730 --> 00:23:24.690 WH11NE- Sunrise: We do something so as an example, we apply this to generation using a particle that graph based classifier for jets how this works is that you take your network. 153 00:23:25.440 --> 00:23:35.880 WH11NE- Sunrise: You feed into a real and generated jets by you take the activations from an ultimate there, and you think out seems to them, and you can appreciate this. 154 00:23:39.000 --> 00:23:54.510 WH11NE- Sunrise: And the point of that you know these high performing pacifiers are basically trying to learn better up to the intuitions about your data in order to do the classification and so, by the end, they should be learning will say lead in futures. 155 00:23:55.770 --> 00:24:03.180 WH11NE- Sunrise: or data to characterize these just so in practice it actually it's it's shown to be very sensitive to the D. 156 00:24:04.320 --> 00:24:07.830 WH11NE- Sunrise: stands for you with a visual inspection. 157 00:24:09.480 --> 00:24:16.380 WH11NE- Sunrise: to your attendees from the one washes scientists proposed it's reproducible and efficient. 158 00:24:17.460 --> 00:24:30.510 WH11NE- Sunrise: It also has more benefits and it's a single aggregate score you can use to evaluate model you retain correlation between all the different features of the race he sees there's no beating. 159 00:24:31.620 --> 00:24:38.490 WH11NE- Sunrise: The only thing missing is the interpretive the ability to the candidate he killed on the score what exactly the model is failing. 160 00:24:39.780 --> 00:24:43.290 WH11NE- Sunrise: To get low score and so it's a square. 161 00:24:44.490 --> 00:24:45.570 WH11NE- Sunrise: conjunction to. 162 00:24:47.070 --> 00:24:48.360 WH11NE- Sunrise: identify the visual. 163 00:24:54.210 --> 00:25:05.610 WH11NE- Sunrise: To which have been pretty effective in practice but i'll also mentioned some alternatives which proposed for that, so a lot of different ways to compare instagrams. 164 00:25:07.470 --> 00:25:11.190 WH11NE- Sunrise: feel divergence Vincent chin average and squares. 165 00:25:12.210 --> 00:25:17.700 WH11NE- Sunrise: And so they're all the reasonable price query particularly this forgiveness offense. 166 00:25:18.510 --> 00:25:35.640 WH11NE- Sunrise: The only benefit for the one washer same distance today takes account of the metrics based so what I mean by this, they can can take this example let's say this is a real tip math distribution and either to distribution predicted by a model and. 167 00:25:36.690 --> 00:25:41.910 WH11NE- Sunrise: If you think about it, the Chi square Jensen killed every. 168 00:25:43.350 --> 00:25:43.770 WH11NE- Sunrise: Every. 169 00:25:45.210 --> 00:25:47.160 WH11NE- Sunrise: Both of these models likes to be the same. 170 00:25:49.320 --> 00:25:53.970 WH11NE- Sunrise: passwords whereas visually whenever you would say that this molecule is. 171 00:25:55.110 --> 00:25:58.320 WH11NE- Sunrise: Clearly, a higher quality better proximity. 172 00:26:00.030 --> 00:26:03.510 WH11NE- Sunrise: But as you see only argument for why i'm saying. 173 00:26:04.650 --> 00:26:05.640 WH11NE- Sunrise: it's not a central. 174 00:26:08.370 --> 00:26:09.300 WH11NE- Sunrise: For completeness. 175 00:26:11.310 --> 00:26:11.820 WH11NE- Sunrise: As much as. 176 00:26:16.230 --> 00:26:16.770 WH11NE- Sunrise: Your question. 177 00:26:20.820 --> 00:26:24.300 WH11NE- Sunrise: Another metric people kind of similar to. 178 00:26:25.500 --> 00:26:42.810 WH11NE- Sunrise: Get the washer same distance classifier it so just use the classifier directly to try to distinguish between really take samples, and if you do this well, can you talk really it's actually very sensitive to the all the new samples and the diversity of samples. 179 00:26:43.980 --> 00:26:52.230 WH11NE- Sunrise: The pros the cons are much everything else, which is a very typical from this, you know from the score you get it. 180 00:26:54.420 --> 00:26:57.720 WH11NE- Sunrise: And it's it's hard to reproduce the same. 181 00:26:58.980 --> 00:27:03.450 WH11NE- Sunrise: Best buyer is not necessarily getting to see it seems or. 182 00:27:05.160 --> 00:27:15.330 WH11NE- Sunrise: it's not particularly fast for the training again as far as the standardized testing everything digitally pacifiers on this on their model. 183 00:27:21.270 --> 00:27:21.870 operation. 184 00:27:23.220 --> 00:27:28.500 WH11NE- Sunrise: So this is why we're building a distance metrics one. 185 00:27:30.120 --> 00:27:30.930 WH11NE- Sunrise: But it's still. 186 00:27:32.640 --> 00:27:33.090 WH11NE- Sunrise: it's all. 187 00:27:35.220 --> 00:27:36.570 WH11NE- Sunrise: come back to this on the last slide. 188 00:27:39.750 --> 00:27:43.890 WH11NE- Sunrise: But i'll move on their new other questions to. 189 00:27:45.180 --> 00:27:47.040 WH11NE- Sunrise: The other Carter dinner love languages. 190 00:27:48.270 --> 00:27:53.790 WH11NE- Sunrise: So in the proposed applications what dimensionality of features are you doing this up. 191 00:27:57.270 --> 00:28:00.630 WH11NE- Sunrise: Each feature coming out in the future, simultaneous do you. 192 00:28:02.790 --> 00:28:03.420 WH11NE- Sunrise: Get the. 193 00:28:04.470 --> 00:28:05.190 WH11NE- Sunrise: Get the generator. 194 00:28:07.140 --> 00:28:15.570 WH11NE- Sunrise: So in So in principle you don't have any to output those features directly, so you actually you know the best. 195 00:28:16.020 --> 00:28:30.600 WH11NE- Sunrise: case would be to output something like output individual bits or particles there just helping the feature article features and then from the samples you calculate the gen X calculates total and you keep track. 196 00:28:31.890 --> 00:28:41.010 WH11NE- Sunrise: Of that so let's say create the generator that produces those tickets yeah and now you want to use one of these metrics evaluated that's using the same one. 197 00:28:42.300 --> 00:28:46.440 WH11NE- Sunrise: How many features, do you want to get the one the. 198 00:28:47.460 --> 00:28:58.050 WH11NE- Sunrise: Data model to fit the link to match much personally yeah good question but it's just about how many of you need until you can trust that it's reducing. 199 00:28:59.910 --> 00:29:02.160 WH11NE- Sunrise: Get mentally, so it is. 200 00:29:03.210 --> 00:29:05.190 WH11NE- Sunrise: really talk about that that gets there's like. 201 00:29:07.320 --> 00:29:16.230 WH11NE- Sunrise: It will come up with these so called energy flow polynomials which form of basis for all all the complete basis for all your. 202 00:29:17.730 --> 00:29:25.110 WH11NE- Sunrise: Piracy safe of Sarah wolves So if you don't get enough of those you're basically saying you've learned yet some scripture right. 203 00:29:26.580 --> 00:29:44.670 WH11NE- Sunrise: yeah not sure coming up with the color shall get just as many as as mania that case and then people are other phrase differently, like how many dimensions to things the water centric will work until it becomes too oh complicated. 204 00:29:45.930 --> 00:29:57.360 WH11NE- Sunrise: So that we were using 50,000 samples or Wednesday, and so, if you want to the same number of things that are in 2d that's already in 50,000 Square. 205 00:29:58.740 --> 00:30:06.720 WH11NE- Sunrise: In fact, is when I was trying to use it, or maybe it could only 32 before it wasn't useful. 206 00:30:08.190 --> 00:30:13.350 WH11NE- Sunrise: So yeah there's a thing at a distance is a better way to get. 207 00:30:15.180 --> 00:30:19.830 WH11NE- Sunrise: There actually another thing we're working on that is not tested at all. 208 00:30:20.970 --> 00:30:21.870 WH11NE- Sunrise: But it's like. 209 00:30:23.250 --> 00:30:26.190 WH11NE- Sunrise: trying to do with something like the fish a distance, but on. 210 00:30:28.110 --> 00:30:30.930 WH11NE- Sunrise: Those games like any any pre work. 211 00:30:37.980 --> 00:30:38.130 WH11NE- Sunrise: dancer. 212 00:30:41.370 --> 00:30:43.200 WH11NE- Sunrise: yeah it's all it's all that talk about. 213 00:30:44.280 --> 00:30:47.160 WH11NE- Sunrise: either for the occasion so like, how do we react to table to. 214 00:30:49.260 --> 00:30:59.970 WH11NE- Sunrise: apply these different generated locals um yeah so it's a little tricky sometimes because the data, you get at the other words it's not. 215 00:31:01.560 --> 00:31:09.390 WH11NE- Sunrise: necessarily similar to typical data and machine learning applications So these are some example with the. 216 00:31:11.460 --> 00:31:14.850 WH11NE- Sunrise: January it's in the detector articles. 217 00:31:16.980 --> 00:31:21.630 WH11NE- Sunrise: And so you can see their data tends to be very sparse highly granular. 218 00:31:23.400 --> 00:31:33.510 WH11NE- Sunrise: possessing potentially or your geometry, and you know the one way to do it is to just kind of take all the data type of form it. 219 00:31:34.380 --> 00:31:45.810 WH11NE- Sunrise: into looking at the image countermeasure shower get image and just this is mostly for they can use convolution neural networks, which have been very popular in. 220 00:31:46.950 --> 00:31:47.310 WH11NE- Sunrise: varying. 221 00:31:48.540 --> 00:31:49.530 WH11NE- Sunrise: varying advance. 222 00:31:50.580 --> 00:31:58.620 WH11NE- Sunrise: But it may not necessarily be the best representation or data losing the brain Larry box let. 223 00:31:59.790 --> 00:32:05.580 WH11NE- Sunrise: me figure pixels it's actually quite hard to train a network to produce the very sparse images. 224 00:32:06.960 --> 00:32:18.030 WH11NE- Sunrise: And you know it's it's inefficient to see all that space so insanely good, for example, just work for me with the kids are the particles save their teachers and he. 225 00:32:19.110 --> 00:32:25.140 WH11NE- Sunrise: You know list, which is done some times, and that is sparse and regular. 226 00:32:27.000 --> 00:32:33.600 WH11NE- Sunrise: you're losing something else which is or you're trying to implement bordering on data which didn't hear the word. 227 00:32:35.580 --> 00:32:36.210 WH11NE- Sunrise: instead. 228 00:32:37.500 --> 00:32:39.900 WH11NE- Sunrise: What we're proposing to. 229 00:32:41.010 --> 00:32:46.890 WH11NE- Sunrise: do now is to use point pounds, which is not only that case or the articles. 230 00:32:48.090 --> 00:32:50.340 WH11NE- Sunrise: or country decision. 231 00:32:51.720 --> 00:32:52.020 letter. 232 00:32:53.070 --> 00:32:55.500 WH11NE- Sunrise: Others, flexible and work for any underlying young. 233 00:32:57.000 --> 00:33:05.250 WH11NE- Sunrise: They have it's just it's better points that there hasn't worked so What this means is that you take the it's in the park and he represented them. 234 00:33:06.360 --> 00:33:12.360 WH11NE- Sunrise: save them in some physical space and you keep their all the features that you want in. 235 00:33:13.380 --> 00:33:13.800 This. 236 00:33:15.810 --> 00:33:19.710 WH11NE- Sunrise: And it's a it's a natural representation for data. 237 00:33:21.300 --> 00:33:34.440 WH11NE- Sunrise: orthopedic motivated character, the geometry aspects which are you and what we found is that with this data, structure and pastoral networks which exploited the typical information and. 238 00:33:35.190 --> 00:33:46.530 WH11NE- Sunrise: interaction between the nodes you guys stayed up your results and see that's the One example is this article heads around this works for who subject headings. 239 00:33:47.610 --> 00:33:48.000 WH11NE- Sunrise: Having. 240 00:33:49.800 --> 00:33:50.220 WH11NE- Sunrise: The. 241 00:33:51.330 --> 00:33:54.450 WH11NE- Sunrise: Learning for part of the flow, which uses words per article. 242 00:33:56.880 --> 00:34:02.910 WH11NE- Sunrise: So i'll just leave this section with the question we extend you know, the state of your work, but to generated. 243 00:34:06.090 --> 00:34:10.890 WH11NE- Sunrise: So, before I move on to actual model development. 244 00:34:18.600 --> 00:34:19.260 WH11NE- Sunrise: So. 245 00:34:20.580 --> 00:34:39.450 WH11NE- Sunrise: So I should say that there's some current application there's been a lot of work actually on on gender modeling so all of you, but if I missed anybody let's move on to the one kind of somewhat famous example, which is a fast how again is used as part of the. 246 00:34:40.590 --> 00:34:44.790 WH11NE- Sunrise: Fast simulation at atlas and it's already we use actually your. 247 00:34:46.950 --> 00:34:50.850 WH11NE- Sunrise: First house innovations 7 billion events already for them to announce. 248 00:34:52.320 --> 00:34:56.310 WH11NE- Sunrise: It technically to conditional washers saying again. 249 00:34:57.690 --> 00:34:59.220 WH11NE- Sunrise: The January shower images. 250 00:35:00.270 --> 00:35:02.790 WH11NE- Sunrise: It says reasonable performance. 251 00:35:04.410 --> 00:35:07.380 WH11NE- Sunrise: They have but really there's movement. 252 00:35:09.180 --> 00:35:17.490 WH11NE- Sunrise: Elsa you did box ization to do the smart city and the hydrating Leonard he's really smart he hasn't got the power editor and. 253 00:35:19.350 --> 00:35:23.040 WH11NE- Sunrise: foxholes which then you're losing the gravy. 254 00:35:24.450 --> 00:35:26.460 WH11NE- Sunrise: And you know it's not clear how that would work. 255 00:35:29.250 --> 00:35:34.230 WH11NE- Sunrise: And there's also training at 300 separate for each data which is pretty pivotal. 256 00:35:36.510 --> 00:35:38.280 WH11NE- Sunrise: Then we'll see how that's. 257 00:35:41.310 --> 00:35:41.850 WH11NE- Sunrise: More than one. 258 00:35:43.530 --> 00:35:47.610 WH11NE- Sunrise: it's like just a way to make it conditional so imagine. 259 00:35:54.990 --> 00:35:57.060 WH11NE- Sunrise: me to use a different set of. 260 00:36:02.580 --> 00:36:05.400 WH11NE- Sunrise: Okay, in a single data, no, no. 261 00:36:06.510 --> 00:36:10.890 WH11NE- Sunrise: Each for each different data, the other games or. 262 00:36:13.020 --> 00:36:14.490 WH11NE- Sunrise: yeah yeah. 263 00:36:18.060 --> 00:36:28.170 WH11NE- Sunrise: But either energy distributions or like I think this is the total energy related shower for different energy input particles. 264 00:36:34.800 --> 00:36:36.000 WH11NE- Sunrise: Is this is total. 265 00:36:37.350 --> 00:36:41.400 WH11NE- Sunrise: And she can do to the level of Americans. 266 00:36:43.230 --> 00:36:46.530 WH11NE- Sunrise: yeah well not not only to the box. 267 00:36:49.230 --> 00:36:56.130 WH11NE- Sunrise: yeah the only to the create they don't do records they like converter olds that are kids into a single. 268 00:36:57.630 --> 00:37:00.030 WH11NE- Sunrise: cylinder niche yeah you. 269 00:37:01.680 --> 00:37:02.430 WH11NE- Sunrise: do get into that. 270 00:37:04.470 --> 00:37:04.680 WH11NE- Sunrise: Is. 271 00:37:06.930 --> 00:37:09.030 WH11NE- Sunrise: super important but which one is. 272 00:37:11.070 --> 00:37:15.090 WH11NE- Sunrise: Well, so like to read in the book oh I. 273 00:37:19.320 --> 00:37:22.770 WH11NE- Sunrise: don't remember hundred percent yeah it doesn't really matter. 274 00:37:24.720 --> 00:37:26.190 WH11NE- Sunrise: I think the red is the gap. 275 00:37:33.630 --> 00:37:40.800 WH11NE- Sunrise: defendant gallons per minute beach sorry i'm just trying to one of getting out of the government workforce. 276 00:37:41.850 --> 00:37:45.360 WH11NE- Sunrise: are predicting out you take them into. 277 00:37:48.900 --> 00:37:49.410 WH11NE- Sunrise: The shower. 278 00:37:54.720 --> 00:37:58.920 WH11NE- Sunrise: What distribution so when you have like a Gavin he he got five. 279 00:38:00.990 --> 00:38:04.320 WH11NE- Sunrise: What are you, what are the time to some degree. 280 00:38:05.820 --> 00:38:09.090 WH11NE- Sunrise: it's just sugar shipping power yeah. 281 00:38:12.900 --> 00:38:16.920 WH11NE- Sunrise: But there's no relationship with one Gal that are completely independent. 282 00:38:19.800 --> 00:38:22.770 WH11NE- Sunrise: Of what are the inputs here just configured on particles after. 283 00:38:24.660 --> 00:38:31.200 WH11NE- Sunrise: yeah like to play on at a particular veto some energy right it's conditional on the end of the incorrect partial. 284 00:38:32.940 --> 00:38:34.710 WH11NE- Sunrise: So we can invite like free time after. 285 00:38:37.320 --> 00:38:43.500 WH11NE- Sunrise: It is this the default our staff or no it's just fess simulation or something. 286 00:38:44.580 --> 00:38:44.940 Okay. 287 00:38:47.820 --> 00:38:53.010 WH11NE- Sunrise: it's also only one part of this whole fast and fire so it's only for tiger on. 288 00:38:54.210 --> 00:38:54.480 WH11NE- Sunrise: TV. 289 00:38:55.560 --> 00:38:56.730 WH11NE- Sunrise: They have their own other. 290 00:38:59.310 --> 00:39:06.600 WH11NE- Sunrise: So, so, for example, for for next one last one, and seems like constantly for the collaborators. 291 00:39:08.970 --> 00:39:09.360 WH11NE- Sunrise: On this. 292 00:39:12.360 --> 00:39:15.600 WH11NE- Sunrise: box vacation they have not mine. 293 00:39:16.740 --> 00:39:22.020 WH11NE- Sunrise: Thank you yeah I guess I don't know what what how exactly they're being used. 294 00:39:23.100 --> 00:39:23.550 WH11NE- Sunrise: Right now. 295 00:39:25.650 --> 00:39:27.270 WH11NE- Sunrise: yeah maybe they're only you. 296 00:39:32.130 --> 00:39:32.700 WH11NE- Sunrise: can see. 297 00:39:34.740 --> 00:39:37.500 WH11NE- Sunrise: You remember how much faster, this is compatible so. 298 00:39:38.850 --> 00:39:55.470 WH11NE- Sunrise: That was my thing they put like all compared to which, like other nations all john so compared to Japanese i'd like the low morale fasten will be 20% of the cpu needed for the full again I don't know, can you tell her. 299 00:39:59.850 --> 00:40:00.210 WH11NE- Sunrise: yeah. 300 00:40:01.680 --> 00:40:03.900 WH11NE- Sunrise: She honestly krista compare. 301 00:40:06.090 --> 00:40:11.760 WH11NE- Sunrise: How do you how do you decide where this performance is acceptable or not really. 302 00:40:13.170 --> 00:40:17.640 WH11NE- Sunrise: yeah that's a good question that's something which would basically have to decide this. 303 00:40:19.290 --> 00:40:24.150 WH11NE- Sunrise: experiment basically gone as far as I know, they didn't have like another. 304 00:40:26.310 --> 00:40:37.710 WH11NE- Sunrise: cms lesson, so this one, just like better than anything else, like which was fascinated yeah so yeah hopefully we'll be able to different models and. 305 00:40:39.000 --> 00:40:40.830 WH11NE- Sunrise: and try to set a threshold for. 306 00:40:42.480 --> 00:40:44.040 WH11NE- Sunrise: Certain level agreement. 307 00:40:45.840 --> 00:40:49.140 WH11NE- Sunrise: Because, presumably, there will be a point be. 308 00:40:51.210 --> 00:40:53.940 WH11NE- Sunrise: Systematic from the fast. 309 00:40:56.220 --> 00:41:07.740 WH11NE- Sunrise: So the same sizes your statistical areas that you're trying to cut down the first reason that point some does it make sense to create the same thoughts for like. 310 00:41:10.380 --> 00:41:17.790 WH11NE- Sunrise: Two samples simulated from the full simulator and and seeing how this compares. 311 00:41:18.930 --> 00:41:32.100 WH11NE- Sunrise: yeah exactly that's that's kind of the bootstrapping thing I was mentioning that you could use was happening between just samples from the real and get like a baseline score in that sense, and we do that or simulations. 312 00:41:34.200 --> 00:41:36.660 WH11NE- Sunrise: that's in a space like that, if it's within. 313 00:41:38.070 --> 00:41:38.430 WH11NE- Sunrise: Those. 314 00:41:46.080 --> 00:41:53.820 WH11NE- Sunrise: Another quick example is big team that that founded information bottleneck and. 315 00:41:55.380 --> 00:41:57.630 WH11NE- Sunrise: They basically use a D in. 316 00:41:59.970 --> 00:42:08.910 WH11NE- Sunrise: And they do this for for time island charm it is, it seems to be some agreement is assembled by cms. 317 00:42:10.470 --> 00:42:11.520 WH11NE- Sunrise: simulations. 318 00:42:14.670 --> 00:42:17.520 WH11NE- Sunrise: Hello flo which is you normalize and close for. 319 00:42:18.600 --> 00:42:27.900 WH11NE- Sunrise: For shower images like Johnson and shower mention again good to see agreement with that looks like because simulations. 320 00:42:29.700 --> 00:42:39.300 WH11NE- Sunrise: three or more sparse approaches, so this is one where it's actually going to 10 articles Eric feature articles modeling the detector. 321 00:42:41.190 --> 00:42:45.390 WH11NE- Sunrise: effect on the particles, but it's just storing the particles over. 322 00:42:47.940 --> 00:42:49.170 WH11NE- Sunrise: To do this with the D. 323 00:42:50.790 --> 00:42:56.520 WH11NE- Sunrise: Is kind of similar by using close as the prayer therapy slightly better tired. 324 00:42:59.250 --> 00:43:11.490 WH11NE- Sunrise: What is another interesting one is analysis is fasting proposal where they actually go directly to a level features from the from tenable. 325 00:43:12.570 --> 00:43:13.110 WH11NE- Sunrise: directly. 326 00:43:14.910 --> 00:43:19.350 WH11NE- Sunrise: He would need an analysis, so this is um it's very interesting. 327 00:43:21.330 --> 00:43:21.720 WH11NE- Sunrise: we'll. 328 00:43:24.330 --> 00:43:27.960 WH11NE- Sunrise: See how these approaches it's it's all. 329 00:43:31.980 --> 00:43:38.880 WH11NE- Sunrise: One drawback, the most a phase of that they do the base representation, as I mentioned it's. 330 00:43:40.410 --> 00:43:49.980 WH11NE- Sunrise: All this hard to scale to a chill senior care to buy something as granular as you should tell them. 331 00:43:51.090 --> 00:43:53.550 WH11NE- Sunrise: And so, for my. 332 00:43:54.780 --> 00:43:55.230 WH11NE- Sunrise: photo. 333 00:43:56.610 --> 00:43:58.080 WH11NE- Sunrise: graph based approach. 334 00:43:59.490 --> 00:44:02.310 WH11NE- Sunrise: called mtn and the key ideas here is having more. 335 00:44:03.540 --> 00:44:11.280 WH11NE- Sunrise: representation for data, and also because we're using a craft we're able to learn global features, as well as your article. 336 00:44:13.110 --> 00:44:15.630 WH11NE- Sunrise: Earlier, since we're trying to learn some structures. 337 00:44:17.580 --> 00:44:22.680 WH11NE- Sunrise: so quickly go over the architecture, we developed for this will use again. 338 00:44:23.730 --> 00:44:24.750 The best performance. 339 00:44:25.980 --> 00:44:26.670 WH11NE- Sunrise: Far and. 340 00:44:28.410 --> 00:44:32.430 WH11NE- Sunrise: what's called a message passing neural network and work for the January. 341 00:44:35.010 --> 00:44:39.990 WH11NE- Sunrise: So the generator were to follow it started off with just noise, for all the features. 342 00:44:41.610 --> 00:44:46.650 WH11NE- Sunrise: and legal itself and what iteration and message testing consists of each node. 343 00:44:48.690 --> 00:44:56.700 WH11NE- Sunrise: getting information all its various forms of learning function F G and aggregating that information. 344 00:44:59.220 --> 00:45:05.970 WH11NE- Sunrise: Is this was done in parallel, for all the notes for however many variations of your hopefully operating. 345 00:45:07.710 --> 00:45:08.310 WH11NE- Sunrise: cloud. 346 00:45:09.510 --> 00:45:21.570 WH11NE- Sunrise: If you're a final teachers and the discriminated takes the generated and we're real power again go through message passing gets an intermediate representation and us into another say what it thinks that. 347 00:45:23.790 --> 00:45:42.360 WH11NE- Sunrise: This is the architecture, the data said it's just we use this to test then just keep it like Bob the variable like yeah I think it's more yet so there's yeah it can be it's just like using masking to mask away. 348 00:45:44.970 --> 00:45:52.380 WH11NE- Sunrise: But smaller sites out on the fake ones and the generators or so, then I guess so. 349 00:45:53.790 --> 00:45:54.450 WH11NE- Sunrise: ignore those. 350 00:45:56.700 --> 00:45:56.880 WH11NE- Sunrise: As. 351 00:45:58.980 --> 00:46:07.110 WH11NE- Sunrise: Well, as lightning but for them it, yes, even Max the mask out certain particles yeah that's a big thing we can develop. 352 00:46:09.720 --> 00:46:12.450 WH11NE- Sunrise: So also true construction site. 353 00:46:17.310 --> 00:46:28.290 WH11NE- Sunrise: I think so they're using another framework and you're able to actually just do message passing on variables that it's about they're actually the young guys in Harry they don't need any. 354 00:46:29.340 --> 00:46:31.530 WH11NE- Sunrise: This was just our way of doing it. 355 00:46:36.480 --> 00:46:44.190 WH11NE- Sunrise: For regenerative models in the US reconstruction, yes, is it is it like as a technology. 356 00:46:45.480 --> 00:46:49.710 WH11NE- Sunrise: Partner question I guess dances yes yeah but it's different. 357 00:46:53.940 --> 00:46:56.970 WH11NE- Sunrise: i'm sorry one thing we're just a great. 358 00:47:00.300 --> 00:47:10.200 WH11NE- Sunrise: Those into the test bench it gets really the library take the 30 highest particles and again like to do some masking. 359 00:47:12.180 --> 00:47:16.740 WH11NE- Sunrise: On between their own you take the relative ITT features. 360 00:47:18.000 --> 00:47:19.980 WH11NE- Sunrise: features which are predicting for carnival. 361 00:47:21.930 --> 00:47:27.870 WH11NE- Sunrise: is going from Danny particles reconstructed gentle while you're showering detector. 362 00:47:29.280 --> 00:47:33.510 WH11NE- Sunrise: And we basically 17 and it took was like working for. 363 00:47:34.860 --> 00:47:45.240 WH11NE- Sunrise: You so let's go over the results on on the top works So these are some sample future you start with 30 articles and then you get the variable number of reconstructive. 364 00:47:48.150 --> 00:47:56.400 WH11NE- Sunrise: Know we're kind of there's no start with that you're just trying to get you started with a single going. 365 00:47:58.680 --> 00:48:13.050 WH11NE- Sunrise: Oh you're saying you Jenner you generate 30 particles of your data settings so you needed a data set to try to reproduce yeah so the original data set has a variable number of articles by a Max of 3030. 366 00:48:14.490 --> 00:48:28.980 WH11NE- Sunrise: That makes sense okay again you MAC so the Max 30 yeah and then, when we generate we're just trying to reproduce that data set, but the input to the generator that just a single article, we would get like a cool. 367 00:48:32.250 --> 00:48:40.110 WH11NE- Sunrise: Did you play it's just so cool always check yeah so also does a product short yeah it does a shower. 368 00:48:41.310 --> 00:48:48.690 WH11NE- Sunrise: Do you see the output is a jet what does that mean the output is them or been reconstructed for backers yeah. 369 00:48:53.220 --> 00:48:54.630 WH11NE- Sunrise: So the nice thing. 370 00:48:55.770 --> 00:49:00.600 WH11NE- Sunrise: Is that, and so we can take their individual articles features, such as the article or else. 371 00:49:01.830 --> 00:49:13.110 WH11NE- Sunrise: We can take jet level features and bash, and this is an example of an energy for polynomial whichever side it's the fps or would complete basis or all. 372 00:49:14.700 --> 00:49:15.870 WH11NE- Sunrise: Get hired cj. 373 00:49:17.370 --> 00:49:21.720 WH11NE- Sunrise: Actually, safe and guess what we're looking at individual projects. 374 00:49:22.980 --> 00:49:32.190 WH11NE- Sunrise: We should be pretty confident it's learning, such as well, that is not one here, so it gives you the real and dread are generated one. 375 00:49:32.760 --> 00:49:47.640 WH11NE- Sunrise: ooh and we also compare it to a fine models, with just buying power generators from machine learning divisions in town, so we can see visually like the best years of other model but also. 376 00:49:48.840 --> 00:49:49.500 The ones. 377 00:49:52.470 --> 00:50:01.230 WH11NE- Sunrise: These are our quantitative scores on the different metrics I mentioned so taking the particle or the math of people support. 378 00:50:03.600 --> 00:50:19.800 WH11NE- Sunrise: Article that distance in all the data model s I mean what, what are the Jacks speak on starting from your tradition was like the features as well. 379 00:50:20.820 --> 00:50:27.960 WH11NE- Sunrise: No, I mean just definition yet oh yeah we're just using so we're trying to create everything we get. 380 00:50:30.240 --> 00:50:30.750 WH11NE- Sunrise: edited. 381 00:50:36.180 --> 00:50:37.410 WH11NE- Sunrise: or something like simplified. 382 00:50:43.110 --> 00:50:43.440 WH11NE- Sunrise: version. 383 00:50:44.850 --> 00:50:57.900 WH11NE- Sunrise: Okay, so it makes a good shot so like have i'm sorry to hear what your plans are the truth level inputs to what you're comparing your simulation to use a blue on and then book in my. 384 00:50:58.950 --> 00:51:11.700 WH11NE- Sunrise: Previous point for Colin around it and then figure out what all the technical partners are, and then I my output is again, I have my q1 and I put particles in a pool around my. 385 00:51:14.100 --> 00:51:21.450 WH11NE- Sunrise: So the that, so there are two different individually, you should distinguish like the the truth data set. 386 00:51:21.930 --> 00:51:25.800 WH11NE- Sunrise: Which we're trying to learn these morons and then there's the input to the generator with you're. 387 00:51:26.400 --> 00:51:41.040 WH11NE- Sunrise: Trying to generate sorry what the truth is like how we define that word yes yeah yeah so that is we just like do a single particle are showering like some detector effects, and then the particle which come out with it and get there there's like a cave. 388 00:51:42.330 --> 00:51:48.150 WH11NE- Sunrise: So you're looking in a colon of size eight around your initial yeah. 389 00:51:50.070 --> 00:51:58.200 WH11NE- Sunrise: So you can think of what the time the chapters and he's like discrete objects, I mean obviously there's no going to be they're not going to clear connection of things and all the things. 390 00:51:59.220 --> 00:52:00.000 WH11NE- Sunrise: But you're. 391 00:52:01.260 --> 00:52:02.760 WH11NE- Sunrise: you're defining a size. 392 00:52:03.780 --> 00:52:06.810 WH11NE- Sunrise: Like, by definition, I guess you're always going on with that yeah. 393 00:52:09.240 --> 00:52:23.010 WH11NE- Sunrise: I should say, this was more like this if there was more for the model development we know like graph just understanding under the stars yeah sure I think a better application would be to get to the to go to records from like a single gene. 394 00:52:25.320 --> 00:52:25.740 WH11NE- Sunrise: is doing. 395 00:52:27.540 --> 00:52:36.390 WH11NE- Sunrise: yeah yeah this is Dr jack sorry you're getting the constructor assigned the real data Center it's matched to the park turn. 396 00:52:36.840 --> 00:52:48.810 WH11NE- Sunrise: That into park the input part of the generator we only put one like this yeah and then it's matched though today jack's to the particles of the data that you are comparing in the real data. 397 00:52:50.550 --> 00:52:59.340 WH11NE- Sunrise: say the glue and then you have any constructive jack yeah they're closer to the data to the particles, and this is JESSICA says matched to the blue ones that you start with. 398 00:53:01.860 --> 00:53:07.170 WH11NE- Sunrise: Like going to school and shirt how's the ground like initialize. 399 00:53:08.970 --> 00:53:10.650 WH11NE- Sunrise: Right, so that this is. 400 00:53:12.330 --> 00:53:12.930 WH11NE- Sunrise: The graph. 401 00:53:14.460 --> 00:53:14.820 WH11NE- Sunrise: Go back. 402 00:53:19.140 --> 00:53:23.640 WH11NE- Sunrise: To the graph is just initialize with a like it's separate from input. 403 00:53:24.690 --> 00:53:29.790 WH11NE- Sunrise: generator that starts with her a graph of regions, noise and we try to transform it to message. 404 00:53:31.020 --> 00:53:35.460 WH11NE- Sunrise: To look like agenda come out if there's something like. 405 00:53:37.050 --> 00:53:49.200 WH11NE- Sunrise: yeah well the edge of that fully connected nodes are yeah but just just for now that's on the third space sample from your second step is get the point yeah. 406 00:53:51.720 --> 00:53:56.340 WH11NE- Sunrise: Then we use graphics on it to do it faster, but if you could just be in. 407 00:53:57.900 --> 00:54:05.220 WH11NE- Sunrise: The future, so the initial good introduced so that's actually not here, though, that would be like a condition. 408 00:54:06.270 --> 00:54:24.630 WH11NE- Sunrise: Right now, so it's always the same, who are in this well it's learning the distribution up a bunch of fall so it'll be anything like that I get to the like the math that all the debt math you're always on some other lighters of the original Article yeah but actually. 409 00:54:25.650 --> 00:54:39.960 WH11NE- Sunrise: find you for to use this we wouldn't need tell us the condition on a single article, so it uses this also you like to aggregate it's like no no so that's why yes. 410 00:54:41.880 --> 00:54:57.750 WH11NE- Sunrise: It we would we would like to actually make it into like a single game, but it's condition like we can input like a specific data and it learns conditional distribution instead of all the different ones one yeah but 200 times bigger yeah. 411 00:55:02.940 --> 00:55:03.510 WH11NE- Sunrise: Which is this. 412 00:55:04.650 --> 00:55:05.670 WH11NE- Sunrise: Valuable to. 413 00:55:09.750 --> 00:55:16.260 WH11NE- Sunrise: The party, but it was more on the distribution of the gates, you have all the articles that all projects. 414 00:55:19.140 --> 00:55:21.810 WH11NE- Sunrise: With individual articles in the region. 415 00:55:24.480 --> 00:55:28.500 WH11NE- Sunrise: In gen the components yeah yeah. 416 00:55:31.980 --> 00:55:32.010 WH11NE- Sunrise: This. 417 00:55:33.600 --> 00:55:42.030 WH11NE- Sunrise: Big interest of you just for the training, you are he gets the 30 particles like I mean there are a part of but you're catching up to 34. 418 00:55:43.380 --> 00:55:53.850 WH11NE- Sunrise: Years yeah and your been in the generator part of it you're just giving it the gen particle is probably the former actor ordering of the things and getting the particles of that. 419 00:55:54.780 --> 00:56:12.540 WH11NE- Sunrise: yeah so 30 partners, I would love for you to describe your scenes it's it's low so again it's just for like developing poker model it's easier to start with, like smaller that's something else which he said he wanted to put like thousand straight down. 420 00:56:14.580 --> 00:56:17.880 WH11NE- Sunrise: To from this to some type of thing or just. 421 00:56:19.920 --> 00:56:20.670 WH11NE- Sunrise: make my craft. 422 00:56:23.040 --> 00:56:43.800 WH11NE- Sunrise: And just go from there are yeah that's that's actually like you know that's something which reach we decide, or you know what would be the best place to put this like what step, I think we get like from minecraft or even from just can particles and then just simulate. 423 00:56:46.770 --> 00:56:55.890 WH11NE- Sunrise: that's that's already in this case you're simulating the shower and, in this case it's it's also the shame yeah sorry, did you just get the background yeah. 424 00:57:01.980 --> 00:57:13.080 WH11NE- Sunrise: Last layer of this like from noise to something I suppose have some number of nodes and you have to decide if you're like this is going to be a particle or. 425 00:57:14.250 --> 00:57:25.380 WH11NE- Sunrise: Like do you just always use all of them are some features not it's just ah let's just say this one is not particle so yeah I actually tried a bunch of ways of trying to learn whether or not. 426 00:57:26.220 --> 00:57:32.490 WH11NE- Sunrise: This is a particle or not it's like all the different way that learning that math yeah I know you did we send it through. 427 00:57:33.000 --> 00:57:46.170 WH11NE- Sunrise: To get into multiplicity of berries yeah yeah and it's it's like it's really yeah it's difficult and so, in the end we just tell it to produce a set number of all the logs real distribution. 428 00:57:47.910 --> 00:57:55.560 WH11NE- Sunrise: Does that mean effectively it's always 30 particles, well, we can think like 20 or 25, but it has to be completed or. 429 00:57:57.480 --> 00:58:15.750 WH11NE- Sunrise: yeah it's kind of like conditioning on the number of people and it's not like a bad thing right because, like the number of protons to some extent is not an observable right, we only see color emitter deposits, which is a finite volume so it's probably fair actually. 430 00:58:17.640 --> 00:58:24.210 WH11NE- Sunrise: charge okay yeah trust you got me all those parents will point out that still not irc yeah. 431 00:58:26.250 --> 00:58:32.940 WH11NE- Sunrise: So what if you feel like you said aka for that sometimes it will give you some jets one or more. 432 00:58:34.800 --> 00:58:36.810 WH11NE- Sunrise: Can you handle that how do. 433 00:58:38.880 --> 00:58:50.520 WH11NE- Sunrise: You train on it for jets are I mean I mean sometimes you can get to jeff's promising to forget that yeah I agree in practice I don't think this is a very. 434 00:58:51.030 --> 00:59:01.080 WH11NE- Sunrise: busy you can directly use this like to news simulations like unless you're specifically looking for specific need making one Dennis. 435 00:59:01.740 --> 00:59:11.580 WH11NE- Sunrise: Thank goodness why I think a better applications to try to go for rickets then from there, you can just apply standard reconstruction you don't have to do with. 436 00:59:13.680 --> 00:59:16.320 WH11NE- Sunrise: It I don't know what the hell, you want to paint a generator. 437 00:59:21.630 --> 00:59:25.680 WH11NE- Sunrise: You felt like an area of interest right and eventually obsessive did the whole detector. 438 00:59:28.620 --> 00:59:42.480 WH11NE- Sunrise: Essentially, about a particle or the charge it's like you know jets of interest to us and later jets from that have like modeling part of the cases etc head around us, and none of these jet. 439 00:59:45.120 --> 00:59:49.650 WH11NE- Sunrise: Energy flow polynomials capture that kind of effect, I didn't catch your secondary criticism. 440 00:59:51.420 --> 00:59:53.340 WH11NE- Sunrise: So be curious how this. 441 00:59:55.050 --> 00:59:55.470 WH11NE- Sunrise: deals with. 442 00:59:57.300 --> 00:59:58.830 WH11NE- Sunrise: How to deal with this framework. 443 01:00:05.550 --> 01:00:05.910 WH11NE- Sunrise: This is. 444 01:00:19.980 --> 01:00:34.290 WH11NE- Sunrise: I like a I think it's just you avoid all this, if you go to the kids instead if you just take the 10 articles right before they go into this detector unless like you have indicated in. 445 01:00:36.660 --> 01:00:37.380 WH11NE- Sunrise: A very interesting. 446 01:00:38.580 --> 01:00:47.400 WH11NE- Sunrise: Oh, you should have learned that really do if you do a good job Okay, it should learn that sometimes a shower starts a little bit further out. 447 01:00:48.630 --> 01:00:48.870 WH11NE- Sunrise: Okay. 448 01:00:51.300 --> 01:00:55.530 WH11NE- Sunrise: So yeah yeah be proposing like records in the tracker as well. 449 01:00:57.090 --> 01:01:05.520 WH11NE- Sunrise: Okay, well, you can propagate the particles in July, oh so Alex does right like they do a job tracker and then they do their gap after that, so you just propagated. 450 01:01:09.360 --> 01:01:12.840 WH11NE- Sunrise: would be very impressive if it could generate tracks. 451 01:01:14.700 --> 01:01:21.060 WH11NE- Sunrise: That would be me I don't think it's possible i'm sure the thing is that the tracker simulation is not very confrontational intensive actually itself. 452 01:01:37.020 --> 01:01:41.640 WH11NE- Sunrise: Oh yeah let's mention that it's very, we have to be playing score is like I said. 453 01:01:43.050 --> 01:02:04.260 WH11NE- Sunrise: which we can measure of Israel sample if you look at the baseline scores we get from our side, look at her scores for the titles feature comparison to the B side actually within the air, the release from sentence is doing that's going to expect and learning the substructure correctly. 454 01:02:05.550 --> 01:02:16.710 WH11NE- Sunrise: Also results on top deck which are much harder and you have three prong and it also affects both emerging or the clustering so you can end up with my fun merge it with the secondary he. 455 01:02:17.610 --> 01:02:24.570 WH11NE- Sunrise: gets into get features So these are very complex by modal distributions and our models like sale to learn the. 456 01:02:25.950 --> 01:02:26.310 Basic. 457 01:02:27.990 --> 01:02:28.650 WH11NE- Sunrise: clustering. 458 01:02:31.740 --> 01:02:36.600 WH11NE- Sunrise: And these are the scores, and the baseline scores, and again. 459 01:02:40.050 --> 01:02:42.360 WH11NE- Sunrise: w one word or the tears. 460 01:02:47.430 --> 01:02:54.720 WH11NE- Sunrise: going to summarize that deacon our strengths based approach and highly successful if they're very complex physics. 461 01:02:57.000 --> 01:03:06.840 WH11NE- Sunrise: Of find a distributor is actually crucial so for all the downside of like having a network which learns the things we care about the square meter. 462 01:03:08.340 --> 01:03:14.220 WH11NE- Sunrise: It also learning particle correlation with the substructure that actually really useful because it helps the generated. 463 01:03:15.450 --> 01:03:18.630 WH11NE- Sunrise: And it's hard to see coming up with just a function. 464 01:03:19.950 --> 01:03:25.530 WH11NE- Sunrise: analytic loss function which would force this kind of stuff to be learned by there. 465 01:03:26.760 --> 01:03:32.400 WH11NE- Sunrise: And so yeah total and said to extend this to tell the remainder of showers for. 466 01:03:33.420 --> 01:03:39.390 WH11NE- Sunrise: each other, they see this particle to reconstruct it it's needed name this case. 467 01:03:40.590 --> 01:03:43.500 WH11NE- Sunrise: It would he have been ideal application. 468 01:03:44.580 --> 01:03:45.660 WH11NE- Sunrise: Or the. 469 01:03:48.420 --> 01:04:04.950 WH11NE- Sunrise: Next steps which we're working on right now, or conditional generation we've discussed conditioning on the gen particle for vectors and learning about conditional distributions scaling to larger thousand back to the thousands of its you would expect. 470 01:04:06.570 --> 01:04:09.090 WH11NE- Sunrise: And just doesn't make any sense for us. 471 01:04:10.140 --> 01:04:11.310 WH11NE- Sunrise: awesome initiatives. 472 01:04:13.710 --> 01:04:21.060 WH11NE- Sunrise: And then finally all just this guy's just leave this slide up for discussion topics for me. 473 01:04:22.500 --> 01:04:24.240 WH11NE- Sunrise: to generate discussion, but. 474 01:04:25.410 --> 01:04:30.870 WH11NE- Sunrise: yeah some summary which is significant opportunity right now to accelerate innovation for each. 475 01:04:32.910 --> 01:04:34.860 WH11NE- Sunrise: The very active research. 476 01:04:36.690 --> 01:04:48.540 WH11NE- Sunrise: doing it, but they're also open questions are cheering for us to decide so i'll mention three, but when we've already discussed they were in the pipeline with it the most effective. 477 01:04:50.160 --> 01:04:55.950 WH11NE- Sunrise: liable to incorporate machine learning think the particle particles direct it's. 478 01:04:56.970 --> 01:04:58.050 WH11NE- Sunrise: Reasonable trade off. 479 01:04:59.430 --> 01:05:00.150 Simplicity. 480 01:05:01.740 --> 01:05:03.030 WH11NE- Sunrise: Of it also wants to. 481 01:05:05.310 --> 01:05:06.720 WH11NE- Sunrise: explore test. 482 01:05:07.980 --> 01:05:09.300 WH11NE- Sunrise: Which model, this is like. 483 01:05:10.350 --> 01:05:20.610 WH11NE- Sunrise: it's difficult cans are very honest or to be the best a distributed, perhaps it also diffusion just came out this role. 484 01:05:21.930 --> 01:05:23.160 WH11NE- Sunrise: models to try. 485 01:05:25.020 --> 01:05:25.590 WH11NE- Sunrise: It also. 486 01:05:26.640 --> 01:05:29.010 WH11NE- Sunrise: find the most important for. 487 01:05:30.120 --> 01:05:30.600 WH11NE- Sunrise: For being. 488 01:05:31.710 --> 01:05:32.130 WH11NE- Sunrise: So how do. 489 01:05:33.630 --> 01:05:35.580 WH11NE- Sunrise: We promote these two which have been very. 490 01:05:39.000 --> 01:05:41.520 WH11NE- Sunrise: easy to use, basically. 491 01:05:48.690 --> 01:05:49.110 WH11NE- Sunrise: Thank you. 492 01:06:02.340 --> 01:06:02.910 WH11NE- Sunrise: reunion. 493 01:06:06.750 --> 01:06:10.080 WH11NE- Sunrise: So you mentioned going to bracket several times. 494 01:06:11.460 --> 01:06:24.720 WH11NE- Sunrise: And I think this is harder in practice, we think it is because going from a seven hit to a record involves the detector electronics, which have conditions that they change. 495 01:06:25.380 --> 01:06:36.780 WH11NE- Sunrise: rapidly, especially the high last year, where you have a large gradient the review, then they also involve things like pile of modeling and may involve calibrations. 496 01:06:37.380 --> 01:06:50.220 WH11NE- Sunrise: So if these things are changing and data on you know, potentially, even a section basis, how will a signal Ghana or other extra vehicle so caps for these different conditions to model that. 497 01:06:51.600 --> 01:06:52.110 WH11NE- Sunrise: Yes. 498 01:06:53.130 --> 01:06:58.980 WH11NE- Sunrise: very good point in different would have to see how rapidly like what is the time to. 499 01:07:01.290 --> 01:07:04.770 WH11NE- Sunrise: richest man in reaching it everything, for example. 500 01:07:07.260 --> 01:07:07.800 WH11NE- Sunrise: The world. 501 01:07:11.700 --> 01:07:30.060 WH11NE- Sunrise: I remember my implication is that if you stop at syndicates you remove a lot of that complexity, I see while still targeted the most computationally intensive okay parts that's it seemed like it was a reasonable it does that happen in modeling portrayed to get media. 502 01:07:31.590 --> 01:07:33.480 WH11NE- Sunrise: I suspect it will be harder and everything. 503 01:07:34.530 --> 01:07:38.970 WH11NE- Sunrise: But you if you stop and think it should be still be very hard. 504 01:07:42.630 --> 01:07:42.810 WH11NE- Sunrise: But. 505 01:07:44.370 --> 01:07:47.730 WH11NE- Sunrise: solely depend on the job training material the material. 506 01:07:49.650 --> 01:07:52.680 WH11NE- Sunrise: Changes in the material fiercely proud we're trying to stop service innovation. 507 01:07:55.290 --> 01:08:06.180 WH11NE- Sunrise: I mean just just giving that endless endless doing now, they don't even go to kids and we have a much more granular either one good. 508 01:08:07.320 --> 01:08:08.250 WH11NE- Sunrise: it's kind of candy. 509 01:08:14.040 --> 01:08:15.660 WH11NE- Sunrise: Just because of the extreme granularity. 510 01:08:22.950 --> 01:08:23.880 WH11NE- Sunrise: particles from each other. 511 01:08:26.760 --> 01:08:29.550 WH11NE- Sunrise: A prediction for distribution for process. 512 01:08:36.420 --> 01:08:37.560 WH11NE- Sunrise: Through distribution. 513 01:08:46.470 --> 01:08:46.980 WH11NE- Sunrise: have asked. 514 01:08:49.290 --> 01:08:51.990 WH11NE- Sunrise: me if you're not in charge of the House. 515 01:08:53.190 --> 01:09:00.450 WH11NE- Sunrise: We don't have an uncertainty well, so this would have been searching and I think the I you can just take this information as yet. 516 01:09:01.980 --> 01:09:05.910 WH11NE- Sunrise: kind of thing to see massive difference so that's what. 517 01:09:07.470 --> 01:09:08.520 WH11NE- Sunrise: comparably. 518 01:09:09.840 --> 01:09:13.950 WH11NE- Sunrise: completed to like compared to the other way yeah if there's a difference if. 519 01:09:15.300 --> 01:09:16.980 WH11NE- Sunrise: You can just check it that's where. 520 01:09:25.020 --> 01:09:25.680 WH11NE- Sunrise: question. 521 01:09:29.610 --> 01:09:30.240 WH11NE- Sunrise: questions. 522 01:09:37.980 --> 01:09:41.370 WH11NE- Sunrise: Certain that it's a similar absolutely aside to kind of. 523 01:09:44.640 --> 01:09:44.940 WH11NE- Sunrise: Take it. 524 01:09:48.840 --> 01:09:54.030 WH11NE- Sunrise: So yeah it depends what you're trying to please keep it you kind of improvement in. 525 01:09:56.400 --> 01:10:14.010 WH11NE- Sunrise: yeah I hear us is not that expensive computation as much as the jam portion unless you're doing very nice city so so that's why it's not take your could it be worthwhile to try to prove it, but if it's just crispy i'm not sure it's the best out. 526 01:10:16.380 --> 01:10:19.800 WH11NE- Sunrise: There are groups using 10 little shadow.