WEBVTT

1
00:34:47.880 --> 00:34:53.640
Redwood C/D - 6504984451: Can you guys hear it?

2
00:34:56.630 --> 00:34:57.310
Wilson, Tom J.: Thanks.

3
00:35:14.790 --> 00:35:17.060
Redwood C/D - 6504984451: So we're in the bridge.

4
00:35:33.250 --> 00:35:34.830
Redwood C/D - 6504984451: And the echo is the echoes.

5
00:35:35.500 --> 00:35:38.330
Redwood C/D - 6504984451: Just… it's all good on. Oh.

6
00:35:40.330 --> 00:35:51.070
Redwood C/D - 6504984451: Alright, so yeah, let's get started. So… This… this session is… Going out a really short.

7
00:35:51.780 --> 00:35:56.489
Redwood C/D - 6504984451: Introductions to problems, so not, like, whole presentations, just…

8
00:35:56.580 --> 00:36:10.999
Redwood C/D - 6504984451: enough to give context on some interesting or perplexing thing, and then we're gonna talk about potential solutions, ways of looking at the problem to diagnose something interesting. But there basically aren't enough

9
00:36:11.000 --> 00:36:22.629
Redwood C/D - 6504984451: opportunities to discuss null results, and null results often lead to really interesting methodological developments. So, even if it couldn't be published because it's, oh, we didn't observe the thing.

10
00:36:22.690 --> 00:36:30.579
Redwood C/D - 6504984451: It's too noisy in this weird way, that kind of thing. It's definitely worth discussing, and collaborating can solve…

11
00:36:30.770 --> 00:36:32.709
Redwood C/D - 6504984451: There's problems that were otherwise.

12
00:36:33.790 --> 00:36:46.190
Redwood C/D - 6504984451: So I will… I will model the first one to just… these can be really quick. So, the… the early leap, the early alerts, these were from…

13
00:36:46.450 --> 00:36:52.709
Redwood C/D - 6504984451: February-ish, there were… there were real bogus scores, reliability scores.

14
00:36:54.610 --> 00:37:06.970
Redwood C/D - 6504984451: there were more alerts with low reliability scores than were initially anticipated, and planning how the alerts would work. So, I was trying to figure out what is the cause of

15
00:37:07.290 --> 00:37:11.000
Redwood C/D - 6504984451: This very large and unexpected number of

16
00:37:11.260 --> 00:37:15.320
Redwood C/D - 6504984451: Middling to low reliability scores.

17
00:37:15.550 --> 00:37:19.849
Redwood C/D - 6504984451: So… Seeing…

18
00:37:20.230 --> 00:37:39.549
Redwood C/D - 6504984451: are they all just one-ups? That was the hypothesis those other people had, and they're not, so there… there's this huge cluster of, when there… there are lots of points in the light curve, so the same object keeps triggering alerts, but they never have high,

19
00:37:39.930 --> 00:37:47.459
Redwood C/D - 6504984451: reliability scores. So… So, that has to be something like image quality or template quality.

20
00:37:47.630 --> 00:37:54.030
Redwood C/D - 6504984451: like, if the template's bad, every time you look there, it's gonna look different from the template. And so…

21
00:37:54.770 --> 00:37:56.670
Redwood C/D - 6504984451: Is there a way that we can…

22
00:37:58.360 --> 00:38:04.999
Redwood C/D - 6504984451: deal with that there's going to be too many alerts to keep this real bogus model.

23
00:38:05.370 --> 00:38:08.350
Redwood C/D - 6504984451: Should we change the algorithm to…

24
00:38:09.080 --> 00:38:12.890
Redwood C/D - 6504984451: The auto-reject based on, like, the history of alerts that are left.

25
00:38:13.010 --> 00:38:19.349
Redwood C/D - 6504984451: That object, that part of the sky, the templates aren't good, they'll always be like that, and so it'll just keep…

26
00:38:19.460 --> 00:38:23.939
Redwood C/D - 6504984451: Triggering too many, so, that…

27
00:38:24.120 --> 00:38:28.849
Redwood C/D - 6504984451: That's an example of, like, a kind of problem that,

28
00:38:29.070 --> 00:38:42.000
Redwood C/D - 6504984451: Some of the discussion already happened, but this is an opportunity if anyone has thoughts on this, and then we'll go on to other members of the ZCC and the other science collaborations who applied.

29
00:38:42.620 --> 00:38:43.390
Redwood C/D - 6504984451: One shot.

30
00:38:44.860 --> 00:38:50.220
Redwood C/D - 6504984451: Yeah, so, let me just get the…

31
00:38:51.440 --> 00:39:01.499
Redwood C/D - 6504984451: There's, Phil Marshall here. So, we've seen an effect just like this as well, trying to find lensed quasars in the alert stream.

32
00:39:01.850 --> 00:39:04.080
Redwood C/D - 6504984451: And…

33
00:39:04.250 --> 00:39:23.089
Redwood C/D - 6504984451: have concluded that persistence of, of appearance in the alert stream is actually a really good signal for us. It's one of the ways we can reject things that haven't been flagged as bogus, but aren't… aren't lens quasars in the alert stream. But yeah, we've…

34
00:39:23.260 --> 00:39:38.149
Redwood C/D - 6504984451: concluded that persistence is important, but I think we've also seen the problem that you're flagging, which is that there are some, some real astronomical objects that are repeatedly being, appearing in the alert stream

35
00:39:38.400 --> 00:39:47.240
Redwood C/D - 6504984451: with, unimpressive reliability scores, and it looks to me as though they might well be, you know, bad PSF models.

36
00:39:47.240 --> 00:40:00.479
Redwood C/D - 6504984451: That kind of thing. I mean, you'd mentioned bad templates, I think it's in the same category of just problems with the… with the processing that leads to things persisting in the alert stream when we'd really rather they didn't. But I did want to flag that…

37
00:40:00.580 --> 00:40:07.619
Redwood C/D - 6504984451: For the lens quasar search, we're looking for things that are… That, that aren't point-like.

38
00:40:07.670 --> 00:40:19.340
Redwood C/D - 6504984451: But we nevertheless still want to find, and so we're in this grey area where, some Lanza AGN may well be flagged as bogus, even though we don't want them to be.

39
00:40:19.340 --> 00:40:29.979
Redwood C/D - 6504984451: So, in our future, we may well be trying to feed back to Ruben a set of, the same set of simulated Lensed AGN that we've been training on.

40
00:40:30.100 --> 00:40:39.140
Redwood C/D - 6504984451: feed it to Rubin in order to improve the real bogus classifier and not reject lens quasars. So I wanted to flag that up as well as something that we're… that we're working on.

41
00:40:40.650 --> 00:40:44.089
Redwood C/D - 6504984451: Thank you. Yeah, so there was a session

42
00:40:44.800 --> 00:40:49.220
Redwood C/D - 6504984451: Yesterday, that… that talked about the real bogus classifier, and…

43
00:40:49.340 --> 00:40:54.000
Redwood C/D - 6504984451: and different strategies for it, and I didn't even know it was on the table to, like.

44
00:40:54.120 --> 00:41:05.689
Redwood C/D - 6504984451: Can we revise it? Can we give it more training? Try again? Like, is it… yeah, revising it over time? I don't know if that's the kind of thing that has to be written into a plan of, like, yes, after this much time, we'll…

45
00:41:05.940 --> 00:41:09.879
Redwood C/D - 6504984451: Redo it, and it'll be different, and then I'll…

46
00:41:10.660 --> 00:41:16.219
Redwood C/D - 6504984451: You are more likely to know the answer to this process for that.

47
00:41:18.330 --> 00:41:20.669
Redwood C/D - 6504984451: Are you saying the real bogus versions?

48
00:41:20.910 --> 00:41:33.480
Redwood C/D - 6504984451: Like, different versions of… Yeah, the real bogus has continuous versions. Three are already available and described in the MTN377, so they typically will be attached to data releases.

49
00:41:33.650 --> 00:41:43.059
Redwood C/D - 6504984451: Because the real book is not only used to give out the real book… the score to the audience, but it's also used to select sources that we then extract.

50
00:41:43.180 --> 00:42:00.589
Redwood C/D - 6504984451: Because we have many more sources than we expected, at least for now, than if, you know, image quality improves, templates improve, that might no longer be the case. But I think we would realistically have a version of the real bogus that is different for every data release, at least in the first few years, and then

51
00:42:01.160 --> 00:42:06.809
Redwood C/D - 6504984451: In addition, different, yeah, versions could be attached to the AP products.

52
00:42:06.920 --> 00:42:15.820
Redwood C/D - 6504984451: If we have a significant performance improvement, we would probably update the version for the alert production.

53
00:42:15.930 --> 00:42:17.050
Redwood C/D - 6504984451: In real time.

54
00:42:21.020 --> 00:42:26.840
Redwood C/D - 6504984451: Okay, so mine was just, like, an example of this, we'll go into…

55
00:42:30.590 --> 00:42:31.630
Redwood C/D - 6504984451: Take it away.

56
00:42:34.450 --> 00:42:36.519
Redwood C/D - 6504984451: And you're being timed.

57
00:42:37.210 --> 00:42:39.730
Redwood C/D - 6504984451: Pardon? Yeah, you got 5 minutes.

58
00:42:43.440 --> 00:42:52.270
Redwood C/D - 6504984451: Hello everybody, my name is Prince. I am developing AI models for object detection, instant segmentation, and classification.

59
00:42:52.380 --> 00:42:54.880
Redwood C/D - 6504984451: So this is our deep disk framework.

60
00:42:55.050 --> 00:43:08.020
Redwood C/D - 6504984451: It's basically able to take in the DP1 tiles, or any kind of Khpish tiles from astronomical data, and it can give you the classification, so it can tell you if it's a star or a galaxy, it can tell you the bounding boxes, so it'll put a box around the item.

61
00:43:08.020 --> 00:43:19.009
Redwood C/D - 6504984451: And that's, like, the mask head, so you could also get the mask of these things. So this is a sample, like, Scarlett-curated pseudo-ground code that I created, and a deep disk prediction, for example.

62
00:43:20.390 --> 00:43:22.879
Redwood C/D - 6504984451: You guys can glance to it for a second?

63
00:43:24.840 --> 00:43:26.010
Redwood C/D - 6504984451: Tell us what we're looking for.

64
00:43:26.630 --> 00:43:32.050
Redwood C/D - 6504984451: So yeah, these are all the… all the blues are galaxies, and all the signs are stars here.

65
00:43:32.570 --> 00:43:45.059
Redwood C/D - 6504984451: And basically, yeah, like, it's basically able to put a box around all the galaxies, for example, and make the mask around them, and it's able to say that if the box is basically blue, it will be a galaxy.

66
00:43:45.280 --> 00:44:01.199
Redwood C/D - 6504984451: And this is basically something that I've published in my first paper on classification head. Our model reaches near parity with the random first classifier, and since it's basically a scene-level model, it basically looks at the image itself, it doesn't rely on the catalog as much.

67
00:44:02.420 --> 00:44:20.009
Redwood C/D - 6504984451: Currently I'm working on self-supervised models, so here, since I use curated ground truth here, on the left side, this is… because DP1 doesn't have two ground truth, I use a curated ground truth. I'm trying to remove the reliance on this, so I could just directly feed in the images, and it will be able to spit out all the predictions.

68
00:44:20.470 --> 00:44:35.730
Redwood C/D - 6504984451: Now, the problem, the one case that I have, like, a specific null case is that, initially, I worked on full DP1, and I produced these AI models that, would work with the full DP1, but when I… when it went to publishing, I only published on these two fields.

69
00:44:36.430 --> 00:44:48.040
Redwood C/D - 6504984451: So, this is because, of the stellar density being very low here, and the galaxy density being extremely high, and also because VR1 and all the upcoming ones will be more focused towards,

70
00:44:48.040 --> 00:44:59.350
Redwood C/D - 6504984451: the deeper fields, basically. So this is why, like, we focused on this, but… and there's always a trade-off, like, if the model works really well on these fields, it might not work as well on others, basically.

71
00:44:59.610 --> 00:45:01.400
Redwood C/D - 6504984451: So, yeah,

72
00:45:02.140 --> 00:45:10.220
Redwood C/D - 6504984451: That's, like, one of the things… I mean, eventually I have to make a model that will work on all of them, so I would appreciate any comments or any feedback.

73
00:45:10.580 --> 00:45:29.689
Redwood C/D - 6504984451: One example I can give is, because the galaxy density is really high here, and the star density was low, you could use something like star rating, so you could augment the data of stars, just so it sees more and more of stars. And that way you can improve your stellar, like, classification stuff, or, you know, the stellar accuracies.

74
00:45:30.150 --> 00:45:41.429
Redwood C/D - 6504984451: But yeah, like, as you can see, there is different amounts. So, like, for example, this one has near-zero galaxies, and it has a lot of stars. A model trained on this would not do as good on the low galactic one.

75
00:45:41.780 --> 00:45:57.300
Redwood C/D - 6504984451: Initially, when I was doing, I was still getting, like, 50-60%, but accuracy, but yeah, so the main part is that those results are not publishable, because mostly with these AI stuff, like, you're trying to prove that they are really good, and you reach this 80-90% accuracy.

76
00:46:02.770 --> 00:46:14.990
Redwood C/D - 6504984451: Yeah, so I suppose this is a, supervised learning. Sorry. Yeah, so, I suppose this is a supervised learning. So, how did you train, like.

77
00:46:15.240 --> 00:46:25.289
Redwood C/D - 6504984451: So yeah, this is, currently, my first paper is on supervised learning part, and for sure, we need the ground truth annotation for it. So for that, I used scarlet-curated ground truth.

78
00:46:25.290 --> 00:46:41.800
Redwood C/D - 6504984451: For a classification part, to know if it's a star or a galaxy for just a training sample, I use the ref extendedness parameter. Okay. Or, the box… for the boxes and the mask, I use Scarlet, so Scarlet is able to tell you pixel-wise, that all these pixels, for example, are belonging to me.

79
00:46:41.800 --> 00:46:48.299
Redwood C/D - 6504984451: From that pixel, I can look at the maximum width of the pixel, the maximum height of the pixel, and then that would make the box around me.

80
00:46:48.690 --> 00:47:02.789
Redwood C/D - 6504984451: So this is basically the curated pseudo-ground tool that I use to train, and if you look closely here, like, this BrightStar contamination has been one of a major, like, issues with DP1, where, there's the… if there's a bright star, it would kind of, like.

81
00:47:02.900 --> 00:47:10.400
Redwood C/D - 6504984451: it messes up with the scarlet. For example, the Scarlett tries to de-blend, and it would put, like, a lot of boxes around here.

82
00:47:10.690 --> 00:47:23.810
Redwood C/D - 6504984451: So, I came up with some techniques to clean up that, basically, and I trained it on that, and if you see my poster later on, you might have a closer look on it, but it's able to even do better than the pseudo-ground troop that I trained it on.

83
00:47:25.560 --> 00:47:29.660
Redwood C/D - 6504984451: So, one counter question, so how you generate the ground truth?

84
00:47:30.730 --> 00:47:47.249
Redwood C/D - 6504984451: So your model cannot surpass, like, cannot do better than how you generate the ground truth. Now, if generating the ground truth consumes more time, then it makes sense, but if generating the ground truth is, like, pretty instantaneous, then what's the selling point here?

85
00:47:47.730 --> 00:47:50.969
Redwood C/D - 6504984451: Well, one… for one, like,

86
00:47:53.940 --> 00:48:00.200
Redwood C/D - 6504984451: Well, okay, so for creating the ground truth part, Alam says, for creating the ground truth part, like.

87
00:48:00.300 --> 00:48:03.810
Redwood C/D - 6504984451: First of all, there's no ground truth for DP1 itself.

88
00:48:03.810 --> 00:48:27.210
Redwood C/D - 6504984451: So there's no, like, point that it's, like, surpassing what's… so this is just a scarlet, curated suit of ground truth. We don't have a ground truth, and we don't… we can't, like, stop the science there, right? Oh, I don't have annotations, I don't have ground truth, so let's just not work on this project, right? Like, science has to progress, and so I made this, like, curated ground truth in order to train it, and in many of the tiles that I've seen visually, and, even, for example, like.

89
00:48:27.210 --> 00:48:29.040
Redwood C/D - 6504984451: If you see here,

90
00:48:30.750 --> 00:48:43.929
Redwood C/D - 6504984451: Technically, I use DP1, but for example, the unrecognized blend, when I compare it with Hubble Space Telescope, our model is able to bring this 3% lower, for example. I trained on ref extending its parameter.

91
00:48:43.930 --> 00:48:54.700
Redwood C/D - 6504984451: And, that's basically morphology only. If you look at the morphology-only plots for, like, this, for example, you'll see a lot of stars misclassified as galaxies, like, it dips basically here.

92
00:48:54.700 --> 00:49:06.209
Redwood C/D - 6504984451: Because this is, like, the… as you go to the fainter regime, galaxies start looking more like point-like. So this drops here for morphology only, and this will go up like this. So if you look at the…

93
00:49:08.120 --> 00:49:25.449
Redwood C/D - 6504984451: There's a paper by Gato et al, and they have shown that already, that the random forest is able to kind of, like, recover similar things, basically. But yeah, even though I trained it on RefX standardness, I trained it on this curated ground truth, it does actually do better than what I trained it on. Okay, makes sense, yeah.

94
00:49:29.470 --> 00:49:46.300
Redwood C/D - 6504984451: Yeah, this seems to suggest a need for metrics that don't require a ground truth. Exactly, so this should suggest, on metrics that don't require ground truth, so that's my main PhD-like thesis, project. With DP2 and further, I will be using something called Deno V3.

95
00:49:46.300 --> 00:50:06.879
Redwood C/D - 6504984451: Dean OV3 can work on, like, it's a self-supervised learning model, so it relies on much less amount of annotations. So even if, for example, in DP1, I had a decent amount of overlap with the Hubble Space Telescope data, I could use the Hubble Space Telescope data as the ground truth and train it on that, and then apply it. So, yeah, that's something that I'm currently working on.

96
00:50:06.880 --> 00:50:19.790
Redwood C/D - 6504984451: And, yeah, it's, like, the main part. One more thing is that on this… this thing, you could also add, like, the redshift head, so you could also get redshift from our model, and you could also get, like, the key points of these galaxies, like spiral arms and all that.

97
00:50:19.840 --> 00:50:20.850
Redwood C/D - 6504984451: physically.

98
00:50:21.750 --> 00:50:33.519
Redwood C/D - 6504984451: Yeah, but I'm trying to remove this, because, yeah, this is the main thing, that the reliance on the pseudo-ground truth is not the best thing, and eventually, as LSSC data progresses, we will have lesser and lesser ground truth.

99
00:50:33.570 --> 00:50:44.749
Redwood C/D - 6504984451: So, yeah, I'm trying to have this Deno V3 model. This is a meta-AI model, basically, and you could work it with only, like, 5 or 6% of data.

100
00:50:45.610 --> 00:50:55.550
Redwood C/D - 6504984451: even if I find 5% ground truth somewhere in the overlap of full DR1 or DR2, you can use these models to, like, train admins. They'll still do really well.

101
00:50:57.220 --> 00:51:02.679
Redwood C/D - 6504984451: Just a comparison point, like, self-supervised, like, when you have the labels and stuff completely correct.

102
00:51:02.680 --> 00:51:21.100
Redwood C/D - 6504984451: people reach accuracies of 85-ish percent, and even in the self-supervised learning regime, you can reach a nice, like, 75-76% accuracy if you use Dinov3, like, models. So I would… we would trade off that 10% accuracy with actually training it on the nice, like, ground truth that we can get, even if it's, like, 5%.

103
00:51:22.810 --> 00:51:24.119
Redwood C/D - 6504984451: Okay, thank you.

104
00:51:32.870 --> 00:51:48.260
Redwood C/D - 6504984451: There's gonna be 3 presenters now? Are they… They're, like, 3 separate presentations. They are 3 separate… How you should interpret this. One after another, but they're on the same topic. Bryce, are they here?

105
00:51:49.470 --> 00:52:00.370
Redwood C/D - 6504984451: So we will, I guess, advance the slides for you, or should we… are they all online? Yes, Hurum says… Hurum says… Hurum says I'm on Zoom, and… Oh, okay.

106
00:52:00.840 --> 00:52:04.269
Redwood C/D - 6504984451: Try saying something, because we don't hear.

107
00:52:04.450 --> 00:52:05.790
Hurum Maksora Tohfa: Hi, can you hear me?

108
00:52:06.580 --> 00:52:07.200
Redwood C/D - 6504984451: Yes.

109
00:52:13.760 --> 00:52:14.580
Redwood C/D - 6504984451: Thank you.

110
00:52:18.890 --> 00:52:20.340
Redwood C/D - 6504984451: You should be able to share.

111
00:52:37.570 --> 00:52:38.400
Redwood C/D - 6504984451: Alright.

112
00:52:39.390 --> 00:52:40.070
Hurum Maksora Tohfa: Awesome.

113
00:52:40.320 --> 00:52:42.479
Hurum Maksora Tohfa: Can you hear me?

114
00:52:42.950 --> 00:52:43.640
Redwood C/D - 6504984451: Yep.

115
00:52:43.740 --> 00:52:46.530
Hurum Maksora Tohfa: Okay, great. So…

116
00:52:46.920 --> 00:52:47.450
Redwood C/D - 6504984451: We…

117
00:52:47.450 --> 00:53:05.469
Hurum Maksora Tohfa: We are giving a talk about, all the AI methods that, we have tried with Ruben. I'll talk about the first effort that's, AI Donut, just before we get into everything else, to give, an overview.

118
00:53:05.800 --> 00:53:21.160
Hurum Maksora Tohfa: since Rubin has such a wide field of view, small changes in temperature or tilt can cause the telescope to go out of focus just a little bit, so we have to

119
00:53:21.160 --> 00:53:31.579
Hurum Maksora Tohfa: actively correct for it, throughout the night. And, it has a bendy mirror, which has a bunch of pistons behind it, and,

120
00:53:31.700 --> 00:53:36.379
Hurum Maksora Tohfa: We kind of combine those things with, 50 degrees of freedom.

121
00:53:36.620 --> 00:53:44.430
Hurum Maksora Tohfa: And we could use those corrections to adjust for any changes that happens.

122
00:53:44.730 --> 00:54:00.530
Hurum Maksora Tohfa: So, to do this, we take out-of-focus, pictures. The reason for doing out-of-focus pictures is, if you look at the PSF perfect optics and distorted optics, it's not, super…

123
00:54:00.700 --> 00:54:10.580
Hurum Maksora Tohfa: different in a… just, like, by looking at eye, you can't exactly see what actually caused the distortion. So instead, what we do is we take.

124
00:54:10.580 --> 00:54:10.930
Redwood C/D - 6504984451: That's true.

125
00:54:10.930 --> 00:54:29.620
Hurum Maksora Tohfa: images. That way, because of the shadow of the M2 mirror, you have these, donut-looking objects. And then, if you have, certain kind of aberrations, that becomes easier to calculate. So here's, for example, is a pentofoil. Obviously, this is very exaggerated.

126
00:54:29.620 --> 00:54:32.179
Hurum Maksora Tohfa: But, we can see, like, these kind of effects.

127
00:54:33.900 --> 00:54:51.839
Hurum Maksora Tohfa: So, for the objects in the field, we take, these out-of-focus images. You can see on the right, that's what they look like, and then we cut out, these stamps, and then we go from, the focal plane to people

128
00:54:52.160 --> 00:55:11.930
Hurum Maksora Tohfa: plane, and what we are currently doing is, we estimate a set of Zernickies, and then we run a forward modeling and try to see if that matches what we see, and then calculate, and then use those Zernickies to correct for,

129
00:55:12.700 --> 00:55:25.549
Hurum Maksora Tohfa: correct for… correct to get the telescope in focus again. So what we have worked on is we trained, CNN,

130
00:55:25.650 --> 00:55:41.539
Hurum Maksora Tohfa: That learns the, feature space of the donuts, and then we give it the image features, and the focal type, and position, and wavelength, and then,

131
00:55:42.060 --> 00:55:47.749
Hurum Maksora Tohfa: predict the Zernickies instead of doing the forward modeling, which can take a long time.

132
00:55:47.850 --> 00:55:53.750
Hurum Maksora Tohfa: At first, we have done that on the data from April in 2025,

133
00:55:53.780 --> 00:56:08.160
Hurum Maksora Tohfa: And we have tried a few models, with, a few different variations of the model to see what works best. Since we had some, simulated data before.

134
00:56:08.160 --> 00:56:14.400
Hurum Maksora Tohfa: We tried to see if we could transfer learn, but what we end up finding is training on real data works the best.

135
00:56:14.400 --> 00:56:23.250
Hurum Maksora Tohfa: So then we take it, and then we, test it on Sky in September last year. We found that,

136
00:56:23.730 --> 00:56:43.449
Hurum Maksora Tohfa: that AI Donut was able to focus the telescope in much less time, and it was able to also process 5 times more data. And of course, it is extremely fast, and it was also robust to bending and vignetting. So, obviously, we want to, explore what happens, when we,

137
00:56:43.490 --> 00:56:58.049
Hurum Maksora Tohfa: extend the data across the months, because the weather condition and everything changes. Once we did that, we tried some simulated testing, and we see that AI Donnet is able to converge.

138
00:56:58.050 --> 00:57:11.499
Hurum Maksora Tohfa: To the correct Zernikis, and it's also robust to, lens, as, you can see, and also bad, image quality on certain nights.

139
00:57:12.740 --> 00:57:32.029
Hurum Maksora Tohfa: And these are our results on the test set, and it seems to work well on lower-ordered Zernikis, as well as the high-order Zernikis, which is really great. And we are waiting for on-sky testing sometime soon. So yeah, I'll pass it on to Taiwan.

140
00:57:34.020 --> 00:57:38.139
Redwood C/D - 6504984451: We can pause for questions. Yes, yes, yes, that's true.

141
00:57:38.830 --> 00:57:39.360
Redwood C/D - 6504984451: Anyone?

142
00:57:39.360 --> 00:57:39.710
Hurum Maksora Tohfa: Okay.

143
00:57:42.840 --> 00:57:47.890
Redwood C/D - 6504984451: Yeah, there's, like…

144
00:57:59.960 --> 00:58:03.939
Redwood C/D - 6504984451: There are a few slides.

145
00:58:06.590 --> 00:58:07.409
Redwood C/D - 6504984451: It's a lot.

146
00:58:08.430 --> 00:58:14.849
Redwood C/D - 6504984451: I mean, I mean… I'll just skip these two slides. I'm sorry, but,

147
00:58:14.980 --> 00:58:21.890
Redwood C/D - 6504984451: If you can go back to the slides later, because I don't know who… I think they got inserted in out of order, so yeah, go ahead.

148
00:58:27.440 --> 00:58:31.869
Redwood C/D - 6504984451: Yeah, okay, thank you. Oh, okay, hi everyone,

149
00:58:32.090 --> 00:58:43.279
Redwood C/D - 6504984451: My name is Tabian Liu, and I'm a PhD student here at Stanford and SLAC, and I'm happy to be able to share my… also, I'm a part of AI here, so it's about AI for remote.

150
00:58:46.270 --> 00:58:55.080
Redwood C/D - 6504984451: Okay. So, we know from Hammer's talk, the AI don't take the, default images as input, so…

151
00:58:55.260 --> 00:59:00.539
Redwood C/D - 6504984451: So my project is… is kind of a project that trend… that,

152
00:59:01.220 --> 00:59:16.390
Redwood C/D - 6504984451: It's based on the AirDonne method, and we want to transfer it to the science cameras. So how can we get the default images from science cameras? We do kind of a so-called triple exposure. So, the general practice is, firstly, we

153
00:59:16.390 --> 00:59:28.159
Redwood C/D - 6504984451: Def folks the whole, science camera plane to the intrapocal position, take one exposure, so where we can… where we can get the intrapocal donut, and then…

154
00:59:28.280 --> 00:59:37.730
Redwood C/D - 6504984451: extra focal exposure, get the extra focus on it, and then the final, the last infocus exposure. So, through the kind of a,

155
00:59:38.240 --> 00:59:44.790
Redwood C/D - 6504984451: well, triplet exposure, we can get the paired domain images, which can be fed into the AI domain.

156
00:59:47.470 --> 00:59:58.339
Redwood C/D - 6504984451: Okay, so these slides, I want to show some, information about how the neural network is constructed and how we split the data. So, yeah, the neural network is

157
00:59:58.830 --> 01:00:17.760
Redwood C/D - 6504984451: I mean, very similar to the wavefront sensor AI donut, so we… so the both the intrafocal and extra-focal donut will be… will pass through the VESTED 18 encoder, and then the imagery features will be combined with, some metadata, like the…

158
01:00:18.180 --> 01:00:23.959
Redwood C/D - 6504984451: field… the field position, the band, the detector ID, and Raptor IDs.

159
01:00:24.130 --> 01:00:32.249
Redwood C/D - 6504984451: And then those… all those pictures will be fed into a multilateral perception to make predictions. So there are 22…

160
01:00:32.520 --> 01:00:43.670
Redwood C/D - 6504984451: prediction has in total. So 21 of them will predict the Zerniki coefficients, which describes the refund. And there's another one, predicting the

161
01:00:43.820 --> 01:00:53.669
Redwood C/D - 6504984451: the pair blur, which is just an atmospheric scene. So the target of this prediction is just to be in alignment with the Danish

162
01:00:53.860 --> 01:00:56.790
Redwood C/D - 6504984451: Which is our ground truth, Danish output.

163
01:00:57.390 --> 01:01:11.970
Redwood C/D - 6504984451: Okay, and… okay, so this image shows how… how the data is paid. So we use 20 nights for training, and we take kind of an alternating manner. So, one 20 night, one testing night, one 20 night, one testing night, and…

164
01:01:12.100 --> 01:01:13.700
Redwood C/D - 6504984451: Seems like it then.

165
01:01:14.670 --> 01:01:26.220
Redwood C/D - 6504984451: Okay, so we have a… so the training… the… I mean, the amount of training data is around 2 million from the peers, and also 2 million for… for training, 2 million for tests.

166
01:01:29.170 --> 01:01:33.840
Redwood C/D - 6504984451: Well, okay, so these slides want to show somewhere. So, on the left, it's,

167
01:01:34.420 --> 01:01:40.479
Redwood C/D - 6504984451: example of the… our AI donor production. So, to compare it with our…

168
01:01:40.980 --> 01:01:52.300
Redwood C/D - 6504984451: with a Danish label, which is used as our ground truth. And the right, on the right, is kind of an overall performance evaluation. So you see, on the…

169
01:01:52.440 --> 01:02:09.109
Redwood C/D - 6504984451: Other clean… there are very clean donut stems. The AI donut can… can generate very similar results as the all-ground trees. So you see these columns, so we… which are… these columns, which are the comparison of, predictions, they're very close, so…

170
01:02:09.550 --> 01:02:14.610
Redwood C/D - 6504984451: the result, the result looks good, yeah. And under wide, so…

171
01:02:14.850 --> 01:02:19.450
Redwood C/D - 6504984451: Each diagram is one zerith mode, and the x-axis is a…

172
01:02:20.510 --> 01:02:36.809
Redwood C/D - 6504984451: Danish label, the ground truth, and the YYX is, add on the prediction. So, for perfect reproduction, we will expect the Y equal to X, diagonal, straight line. However, it's not that perfect. So, in reality, we see some,

173
01:02:36.970 --> 01:02:40.710
Redwood C/D - 6504984451: Like, these small outlier points, and

174
01:02:40.900 --> 01:02:46.249
Redwood C/D - 6504984451: Also, the kind of, you see, non-zero biases, and also non-zero…

175
01:02:46.820 --> 01:02:56.870
Redwood C/D - 6504984451: I mean, slopes, I mean, the slopes are not 1, so, which means it's not perfect. So…

176
01:02:57.250 --> 01:03:01.309
Redwood C/D - 6504984451: The good thing is that we can still see some, like, clear correlation.

177
01:03:02.070 --> 01:03:12.380
Redwood C/D - 6504984451: Just one sign, yeah. So maybe it's… I mean, it is still useful under some conditions, but we need more training with more data.

178
01:03:12.600 --> 01:03:20.809
Redwood C/D - 6504984451: Probably with more data or more nights covering different part of the space, like ambient temperature, like, different mechanical state.

179
01:03:20.980 --> 01:03:21.860
Redwood C/D - 6504984451: Excellent.

180
01:03:23.260 --> 01:03:30.720
Redwood C/D - 6504984451: Okay, so this is the last slide. We want to talk about some more basic questions. So, why add on full remote?

181
01:03:30.880 --> 01:03:35.109
Redwood C/D - 6504984451: So the first… the first reason is very… is the same as,

182
01:03:35.260 --> 01:03:43.770
Redwood C/D - 6504984451: So we wanted to accelerate the image processing speed. And the second reason, which is…

183
01:03:43.900 --> 01:03:56.340
Redwood C/D - 6504984451: yeah, it's also similar to the add-on-Franc sensors, that it… we think there's some evidence that shows it's more… it's more robust to the image quality, so there are two examples. This one is,

184
01:03:56.530 --> 01:04:04.420
Redwood C/D - 6504984451: So there's… so you see, at the edge, there are the… There is a neighbor, donut.

185
01:04:04.850 --> 01:04:16.969
Redwood C/D - 6504984451: So this is the first example, with kind of a low-quality image, and the second example is that, so low, low signal-to-noise ratio, and there's some… some strange, noise point.

186
01:04:16.970 --> 01:04:24.989
Redwood C/D - 6504984451: this price point. So you see, on both cases, the AI donut outperforms the…

187
01:04:25.220 --> 01:04:37.139
Redwood C/D - 6504984451: initial, like, our traditional algorithm, the Danish algorithm, because these… these features will drag the, output of the Danish very quickly, which is not good.

188
01:04:37.310 --> 01:04:44.999
Redwood C/D - 6504984451: But… But we are still working on some dedicated training for these local quality images to…

189
01:04:45.110 --> 01:04:48.490
Redwood C/D - 6504984451: Makes… makes the preliminary evidence to be a…

190
01:04:49.000 --> 01:04:53.430
Redwood C/D - 6504984451: to be true, it should be more or less. So…

191
01:04:53.690 --> 01:04:56.239
Redwood C/D - 6504984451: Yes, that's all I want to share. Thanks.

192
01:05:03.470 --> 01:05:08.670
Redwood C/D - 6504984451: Oh yeah, I was gonna say, if there are any quick questions while you're… I blocked your eyes.

193
01:05:08.820 --> 01:05:10.570
Redwood C/D - 6504984451: So, walk around.

194
01:05:11.570 --> 01:05:15.739
Redwood C/D - 6504984451: I have a question about those biases.

195
01:05:16.030 --> 01:05:18.179
Redwood C/D - 6504984451: Previous slide, you know what?

196
01:05:18.320 --> 01:05:23.200
Redwood C/D - 6504984451: So, yeah, what… what causes…

197
01:05:23.750 --> 01:05:43.190
Redwood C/D - 6504984451: The reasonable bias? That's not reflected in the next plot of, like, how good is it overall, that's probably… Yeah, yeah, I mean, this is just a single example, but there is kind of statistics. It's very difficult to locate one single reason, because, you know, it's a machine learning.

198
01:05:44.250 --> 01:05:47.440
Redwood C/D - 6504984451: Like, how are you investigating potential reasons?

199
01:05:47.920 --> 01:05:55.700
Redwood C/D - 6504984451: Will… Do you want help investigating the potential? Yes, yes, yes. Well, we'll have to…

200
01:05:57.030 --> 01:06:06.600
Redwood C/D - 6504984451: Actually, the initial version is much… is even worse, much worse. Like, the slope will be compressed to some length.

201
01:06:06.980 --> 01:06:15.919
Redwood C/D - 6504984451: a big population at the center. And, we have to, improve our training method and, improve, like, enlarge our data set.

202
01:06:16.230 --> 01:06:23.510
Redwood C/D - 6504984451: So more training to help fix the virus. Yeah, so, and kind of fixed, so… There's, already happened.

203
01:06:24.150 --> 01:06:27.159
Redwood C/D - 6504984451: fix, fix, I mean, for improvement.

204
01:06:27.270 --> 01:06:33.240
Redwood C/D - 6504984451: With, larger, larger data set, and… Just improvement of training.

205
01:06:33.510 --> 01:06:41.269
Redwood C/D - 6504984451: So… but still, at this stage, I think we have to consider some physical reasons, like what I mentioned, the…

206
01:06:41.480 --> 01:06:49.130
Redwood C/D - 6504984451: particle space. But, like, we only have 20 months. This is very small. So, maybe we can expand it to…

207
01:06:49.970 --> 01:06:51.360
Redwood C/D - 6504984451: more monsters?

208
01:06:51.880 --> 01:06:58.259
Redwood C/D - 6504984451: Or… more, like, optical state, or, I mean, temperature, things like that, so we have…

209
01:06:58.390 --> 01:07:01.459
Redwood C/D - 6504984451: We can expect this can be improved in the future.

210
01:07:02.930 --> 01:07:04.219
Redwood C/D - 6504984451: Thank you. Thanks.

211
01:07:09.240 --> 01:07:22.750
Redwood C/D - 6504984451: All right, so, thanks, Taruan. I'm gonna continue off of the lost amount of work that Hurum and, Taiyun has done already on AI Donuts, and extend it out further with a model that we're calling Tarts.

212
01:07:23.080 --> 01:07:34.999
Redwood C/D - 6504984451: So, so far on our discussion about AI donuts, and also Danish, is the underlying assumption needs to be true, which is that Danish is the optimal algorithm for estimating the wavefronts.

213
01:07:35.000 --> 01:07:52.710
Redwood C/D - 6504984451: And are we exactly sure that's the case, is what we're going to try to look at right now. So, there's some hints that perhaps this is not the best, but a functional algorithm is that if we're to do a simple test of looking at how well does the lowest orders and keep terms correlate to itself.

214
01:07:52.720 --> 01:08:06.549
Redwood C/D - 6504984451: As it evolves through time. And you can see on the top right here, when you look at Danish, at the lowest order left-hand corner, you see that it pretty much immediately de-correlates from itself after, like, two images taken.

215
01:08:06.620 --> 01:08:22.589
Redwood C/D - 6504984451: And you also can look at, perhaps, like, spatial resolute… spatial noise. If you were to look at when you're just observing, just monitoring, not doing active controls, that the highest order terms of the Xernity expansions are quite noisy.

216
01:08:22.590 --> 01:08:29.830
Redwood C/D - 6504984451: And we're suggesting, potentially, that might be a better solution, when I'm hinting at the other models here.

217
01:08:30.160 --> 01:08:35.880
Redwood C/D - 6504984451: And so, with that assumption put in place, how can we potentially improve upon Danish

218
01:08:35.880 --> 01:08:50.409
Redwood C/D - 6504984451: and also AI donuts in the future, extending that further on the work that people have done already. So we're calling this model TARTS, and it stands for a mouthful. It's Triple Stage Alignment Reconstruction Transformer Systems, and that is

219
01:08:50.410 --> 01:09:00.770
Redwood C/D - 6504984451: 100% chat GPT generated, just to fit with the theme of pastries, because donuts, Danish, and not tarts. But yeah, alright, let's continue about this. So…

220
01:09:00.770 --> 01:09:19.520
Redwood C/D - 6504984451: one method for it to potentially do better is, what if you just got better simulations, right? So, as Hurum has mentioned, that Danish has this underlying ray tracing model that's super fast, do all this for ray tracing in real time. But there's also much more expensive ray tracer models out there, for example, MSIP.

221
01:09:19.520 --> 01:09:24.829
Redwood C/D - 6504984451: which takes on the order of hours to run, which is something you definitely can't be doing on Sky in real time.

222
01:09:24.830 --> 01:09:37.399
Redwood C/D - 6504984451: So what if we just emulated that? Something that includes so much more physics, the atmosphere, CCD physics, the underlying optical model, throw the whole kitchen sink at it, see what happens. So that's one thing that we could potentially do better.

223
01:09:37.569 --> 01:09:51.220
Redwood C/D - 6504984451: The other thing is, what if we just had a larger model, okay? So, this part, what we call the WaveNet, is pretty much inheriting off of AI Donut, okay? What if we just extended that with two more models?

224
01:09:51.220 --> 01:10:11.039
Redwood C/D - 6504984451: Okay, so we have one model that does the donut detection, so we're now not just using, like, the 8 donuts that Danish is using, or, like, the 20-ish donuts that, AI Donut is using, but we're using basically everything that we can possibly get out of these corners, right? So this is on the order of maybe 80 donuts, for example, per exposure, per corner.

225
01:10:11.500 --> 01:10:28.510
Redwood C/D - 6504984451: So, individually, these donuts come up with an estimate of their wavefront, cool, but each donut estimates the same wavefront in the end, supposedly. And so, potentially, you can use all of those donuts to get the even better estimate. And so you have an aggregator at the bottom here that gives you a more global estimate here.

226
01:10:28.990 --> 01:10:50.370
Redwood C/D - 6504984451: But you're probably wondering to yourself, Peter, you can't just be training on simulated data, that's probably not gonna work. So how do you incorporate real data into your model without biasing or not having any labels at all? And so this comes in the field of domain adaptation, and there's a method that we can talk about more, just called Derogram, to bridge this sim-to-real gap here.

227
01:10:51.060 --> 01:11:11.569
Redwood C/D - 6504984451: Okay, so what are the results here? When you get larger simulations, bigger models, and more data, do you actually perform any better when you go on Sky and try to test things? So, just like AI Donuts, we also went on Sky, I guess, more recently now. This was July 14th, or something like that, and so here are some results.

228
01:11:11.570 --> 01:11:26.439
Redwood C/D - 6504984451: We ran our model actively controlling in initial alignment, on R-band in a fixed pointing. And what you're looking at here is the evolution of the PSF full attack max, the median of that in arts seconds as it evolves through time.

229
01:11:26.440 --> 01:11:51.099
Redwood C/D - 6504984451: And what we see here is that the red is tarts, and the blue is Danish, and within the same amount of sequences, we do see that we're converging to a lower floor here, which is pretty exciting. But what is even more exciting is that if we were to let this run for even longer, we're able to get sub.6 arcseconds pretty consistently for on the order of maybe 10-ish images, which is pretty much the duration of that test.

230
01:11:51.780 --> 01:12:07.370
Redwood C/D - 6504984451: Here's some pretty pictures of the focal plane here. This is the medium PSS per each detector, and here are images of, you know, the PSS per part of that focal plane. There's, like, red boxes around it.

231
01:12:07.370 --> 01:12:15.620
Redwood C/D - 6504984451: makes it seem like it's doing bad, but it's, like, pretty good. It's, like, 0.55 arcseconds. So… this is looking pretty good.

232
01:12:15.820 --> 01:12:38.210
Redwood C/D - 6504984451: We also can measure the, gradients of the PSF, so we want the focal point to be pretty uniform in performance, and we see that when we do go on Sky? And you see that when we turn on turns with our model, you see basically a step function change in that performance there, which is indicating potentially actual improvement in the performance.

233
01:12:38.230 --> 01:12:42.169
Redwood C/D - 6504984451: I'll just… Craig, credit score to Craig for how she did that analysis for us.

234
01:12:42.980 --> 01:12:55.680
Redwood C/D - 6504984451: Okay, so, so far, things look good, but again, this is a very idealistic scenario. This is one single pointing, sample size of, like, four, right? So, can't really draw conclusive evidence yet.

235
01:12:55.680 --> 01:13:17.570
Redwood C/D - 6504984451: So are we winning? Not yet. But it's showing some real signs of bridging a simulation to a real gap, which is pretty exciting. You know, from Hurum's original analysis, we saw that, for example, you know, simulations might not be the best solution to solving this problem, and that you need to have real data with labels.

236
01:13:17.570 --> 01:13:34.290
Redwood C/D - 6504984451: But this current analysis has shown that, like, we can actually, you know, bridge that gap with simulations and no labels from real data. And, you know, in the field of machine learning, just to us, it's like, this is like the holy grail of bridging sim to real, so that's the exciting part for us, and

237
01:13:34.290 --> 01:13:56.379
Redwood C/D - 6504984451: That gives us confidence to try to maybe, in the next step, to see to extend this to, all the bands, so not just R, and also try to see if we can, in a more realistic survey mode, deliver good performance as well. And we'll get more statistics about, you know, how our performance is sample size greater than 4, hopefully, but…

238
01:13:56.380 --> 01:14:09.050
Redwood C/D - 6504984451: Yeah, summary notes, but yeah, I'm happy to open floor for any questions, and generally, also, Taiyan can also answer questions as well. Just, I had, like, two takes, one… Just a second.

239
01:14:09.770 --> 01:14:28.410
Redwood C/D - 6504984451: I had, like, two things. One, I realized that the other two people were using ResNet18 or something, so that was the CNN-based architecture, and you're using a transformer-based architecture? It's a transformer and a CNN. Okay, so did you, would you say, like, the advantage came mostly because of using the transformer architecture, or…

240
01:14:28.440 --> 01:14:44.230
Redwood C/D - 6504984451: Yeah, so, the transformer… when we did our ablation study of removing this guy, we saw that there were performance degradation in the estimates when ground truth compared to simulation. We didn't burn any on sky time to…

241
01:14:44.230 --> 01:14:59.770
Redwood C/D - 6504984451: tested ablation, but we did see performance improvement. So it's not super substantial, but there were performance improvements by adding this aggregator at the end. Right, okay. And my second thing was, I actually did something very similar, like, I see you have the real data.

242
01:14:59.770 --> 01:15:11.840
Redwood C/D - 6504984451: And because the people who worked before me, they had been training it on simulated data the whole time, so that when LSSTDP1 comes out, we can see how it works on real data. So, when I applied it on real data,

243
01:15:12.130 --> 01:15:13.340
Redwood C/D - 6504984451: Basically.

244
01:15:13.420 --> 01:15:33.340
Redwood C/D - 6504984451: getting it from the normal models to simulated data, simulation, and then to real data, this did exactly the same for me as just taking the base models and directly just applying it with real data. So I would say, like, because I see you're using, like, a bridging between those two. I would say, like, if you try just directly using the real data to train on.

245
01:15:33.370 --> 01:15:44.909
Redwood C/D - 6504984451: it might… you might not need any, like, going through the simulations. So we don't have… we don't… we should not necessarily trust any of the labels that we have for real data. Oh, okay, okay. Yeah, I see that.

246
01:15:45.490 --> 01:15:46.460
Redwood C/D - 6504984451: Sure.

247
01:15:46.600 --> 01:16:03.310
Redwood C/D - 6504984451: we could do that, and that is a lot of the work that Hurum and Taiyuan is doing, which is their training on Danish outputs. We're kind of attacking this problem orthogonally, in the sense that, what if we had something completely different than Danish? Could we still

248
01:16:03.730 --> 01:16:04.790
Redwood C/D - 6504984451: form.

249
01:16:05.300 --> 01:16:10.930
Redwood C/D - 6504984451: similar, or better, or whatever the case is today, or whatever that might be. That makes sense, yeah.

250
01:16:15.380 --> 01:16:18.019
Redwood C/D - 6504984451: And if I can make an observation.

251
01:16:18.140 --> 01:16:27.669
Redwood C/D - 6504984451: So it seems like it's another case where metrics that don't require round proof There's a lot of subterrys.

252
01:16:28.500 --> 01:16:34.489
Redwood C/D - 6504984451: be valuable, so… Yes. Yes. Metrics in different contexts. Yes, exactly.

253
01:16:34.620 --> 01:16:36.713
Redwood C/D - 6504984451: Got it.

254
01:16:37.630 --> 01:16:40.839
Redwood C/D - 6504984451: Are there other questions for all three?

255
01:16:41.010 --> 01:16:42.710
Redwood C/D - 6504984451: Speakers on this topic?

256
01:16:45.820 --> 01:16:48.680
Redwood C/D - 6504984451: Yeah, let's give all three a round of applause.

257
01:16:59.560 --> 01:17:04.680
Redwood C/D - 6504984451: I changed the slides, but… Probably not.

258
01:17:06.760 --> 01:17:10.129
Redwood C/D - 6504984451: This is an excerpt back upside.

259
01:17:12.140 --> 01:17:13.879
Redwood C/D - 6504984451: Great, yes, let's go.

260
01:17:18.830 --> 01:17:29.670
Redwood C/D - 6504984451: Hello, everyone. This is Jay Bhattacharya from Amherst College. I'm a postdoctoral researcher, and I'm working with Dr. Mia De Los Teres, studying, low-mass dwarf galaxies.

261
01:17:29.670 --> 01:17:42.689
Redwood C/D - 6504984451: And the problem that we are, interested in right now is, like, to, to generate a catalog of low redshift dwarf galaxies that can be of,

262
01:17:42.720 --> 01:17:46.139
Redwood C/D - 6504984451: use for, for the, for the,

263
01:17:46.460 --> 01:17:51.799
Redwood C/D - 6504984451: Nomass Galaxy Low Surface Brightness, Galaxy community. So…

264
01:17:52.180 --> 01:18:09.120
Redwood C/D - 6504984451: So, so the problem, is, like, at, say, redshifts less than 0.1, it is very hard to, constrain, like, photometric redshifts and, other methods, like surface brightness fluctuation, I mean…

265
01:18:09.470 --> 01:18:26.989
Redwood C/D - 6504984451: which has been used to measure, like, redshifts at scale for the local volume, which isn't, like, applicable, let's say, when you are, like, looking at 100 megaparsec or something like that. So, so we are trying to, use the

266
01:18:27.030 --> 01:18:38.220
Redwood C/D - 6504984451: use the XSAGA framework, so that was, this, this ResNet-based classifier that was, designed by our collaborator, John Wu.

267
01:18:38.220 --> 01:18:53.209
Redwood C/D - 6504984451: To, to, identify such low redshift candidates, from, from the legacy survey using, spectroscopic redshifts from, from, from the SAGA, survey. So…

268
01:18:53.750 --> 01:18:59.120
Redwood C/D - 6504984451: We are trying to repurpose that and apply that to LSST.

269
01:18:59.200 --> 01:19:15.669
Redwood C/D - 6504984451: And, today, I will, like, give a… I mean, we already… President… me and, yeah, we presented this at the low surface brightness galaxy Session, but I think after DP2, we are also, like, recalibrating as to, like, what's of, like, the…

270
01:19:16.530 --> 01:19:27.460
Redwood C/D - 6504984451: greatest interest here, and, and also looking at some, some, some, limitations of, of, of, of this. So…

271
01:19:27.930 --> 01:19:30.689
Redwood C/D - 6504984451: So right now, what we… what we are…

272
01:19:30.730 --> 01:19:52.440
Redwood C/D - 6504984451: planning with the GP2 release is that to train this on, say, on a spectroscopic catalog of galaxies, typically greater than 10,000 galaxies, and to deliver this catalog. And over here, what you see are these outputs of the original XSaga.

273
01:19:52.440 --> 01:19:57.110
Redwood C/D - 6504984451: to show what Xaga is… does, so Xaga classifies, like.

274
01:19:57.260 --> 01:20:16.060
Redwood C/D - 6504984451: low near, nearby galaxies against more distant high redshift galaxies. So the galaxies in the upper row, these are the high redshift distant galaxies. They have this lower value of this probability, that's, provided by the

275
01:20:16.450 --> 01:20:29.039
Redwood C/D - 6504984451: the CNN, and the low redshift galaxies you see in the bottom, bottom row. So these are the objects of our interest, and, these have a higher probability. So…

276
01:20:29.880 --> 01:20:35.330
Redwood C/D - 6504984451: So yeah, this is what we are trying to do. So this is legacy survey imaging. We want to, like.

277
01:20:36.230 --> 01:20:55.700
Redwood C/D - 6504984451: do this for LSST, because right now it's TP2, but once we have TR1, and as the survey progresses, it's going to get better, and as we cover a larger area of the sky, I think we can, like, generate a very large sample of such low-mass star galaxies.

278
01:20:57.640 --> 01:21:14.319
Redwood C/D - 6504984451: Yeah, so, so, the… right now, what we have done is we have designed this pipeline that can efficiently, like, extract UGRA cutouts in bulk very efficiently, and we also, like, perform, like, preliminary training using, like, XSACA. And,

279
01:21:14.990 --> 01:21:17.239
Redwood C/D - 6504984451: And yeah, so these are some… some…

280
01:21:17.240 --> 01:21:36.970
Redwood C/D - 6504984451: images from DP1 that we use, like, these are on the top row shows the nearby galaxies, and the bottom shows, like, the distant galaxies. So, as you can see, the nearby galaxies are more extended, and I mean, that's what we are, we are trying for the model to learn, and thereby classify these galaxies.

281
01:21:37.050 --> 01:21:54.729
Redwood C/D - 6504984451: So, so yeah, this is… this is the main project that I've been working on. But aside from this, I'm also interested in galaxy morphologies, because, recently, I think two weeks back, I, I, completed this paper, this work that I've been,

282
01:21:55.230 --> 01:22:10.120
Redwood C/D - 6504984451: engaged in, so that's on the non-parametric morphologies of low-mass galaxies, and I'm using multiband photometry. Not just a signal band, but GRIZ. Once again, this is legacy survey.

283
01:22:10.120 --> 01:22:20.330
Redwood C/D - 6504984451: So, so the utility of that is, if we study the non-parametric morphologies, like the GD coefficient, F2G, concentration, asymmetry, smoothness parameters.

284
01:22:20.510 --> 01:22:37.140
Redwood C/D - 6504984451: it has been found that these, parameters… these, measures are, like, can be… have correlations with the physical properties of the galaxy, like stellar mass, star formation rate, so… so there's some, utility here, I think, that

285
01:22:37.240 --> 01:23:00.629
Redwood C/D - 6504984451: can, like, explore further. So the… on the left-hand side, you see this plot from the paper, where we are doing, like, the image segmentation, and then measuring the morphologies using this code called stat morph. And the thing with the low-mass galaxies that you, like, look at here closely is, like, there, sometimes the galaxies are off, like, irregular shape.

286
01:23:00.630 --> 01:23:15.319
Redwood C/D - 6504984451: Like, sometimes there isn't, like, a… I mean, there's, like, star-forming regions, and also there's lots of stuff going on, and for this, the two parameters, asymmetry and smoothness, I think that

287
01:23:15.320 --> 01:23:26.989
Redwood C/D - 6504984451: that will be very useful, and I think we can capture in a greater detail than the legacy survey, with, with LSST. So, so, so that's what I'm, like, interested in to,

288
01:23:27.360 --> 01:23:36.010
Redwood C/D - 6504984451: whether, if we were to, like, use, something like a neural net or something, or we do, like, the non-parametric

289
01:23:36.010 --> 01:23:57.879
Redwood C/D - 6504984451: measurements and do something like dimension energy reduction. Actually, we do a little bit of that in our paper, so, we apply UMAP to see, like, how the… how the parameters are… are distributed in this space. So, so yeah, this is something I'm working on. I think I'm also, like, part of, like, the… the work of the

290
01:23:57.920 --> 01:24:21.730
Redwood C/D - 6504984451: Galaxy Morphology challenge that's, being undertaken by the Galaxy Science, collaboration. So, so yeah, it is an interesting time, and I also want to, like, compare how, like, these measurements, are between, say, Legacy Survey and SST, LTP2. So, so yeah, so this is, some of the interesting…

291
01:24:22.040 --> 01:24:29.530
Redwood C/D - 6504984451: things that I've been doing, and I would like to answer any questions. If you have, that I would like to cut you by as well.

292
01:24:39.890 --> 01:24:44.040
Redwood C/D - 6504984451: Start discussions, yeah.

293
01:24:45.620 --> 01:24:46.320
Redwood C/D - 6504984451: Okay.

294
01:24:46.480 --> 01:24:57.400
Redwood C/D - 6504984451: I'm curious to know, I'm sure I could read the paper, but you can also tell me about how, you, were able to cluster your sources, and which algorithm you used.

295
01:24:57.750 --> 01:25:14.500
Redwood C/D - 6504984451: Is this even on? It's for the people online. Oh, okay, not for the room, so I need to speak louder anyway. Yeah, which, learning method or algorithm did you use to, to create the representation that you've visualized in your UMAP space there?

296
01:25:14.970 --> 01:25:21.830
Redwood C/D - 6504984451: Oh, so, so yeah, I, I had something like 20, sort of, like,

297
01:25:22.060 --> 01:25:26.490
Redwood C/D - 6504984451: Data points for each galaxy, and the number of galaxies were, like, 2,000.

298
01:25:26.620 --> 01:25:39.800
Redwood C/D - 6504984451: So, so yeah, so that's, I mean, a large thing, and I think, I thought, like, UMAP, this, this uniform manifold, some sort of, like, projection, this was, like, one of, like, the,

299
01:25:39.800 --> 01:25:47.919
Redwood C/D - 6504984451: methods that people tend to use for dimensional deduction reduction to see how things are testing. So, this is more of, like, a

300
01:25:48.120 --> 01:25:51.679
Redwood C/D - 6504984451: Proof of concept, rather than the main thing in this paper.

301
01:25:51.820 --> 01:26:02.590
Redwood C/D - 6504984451: But, I think, I know, like, other groups, they are also, like, considering, like, methods like climate change reduction to, like, study, like, galactic morphologies.

302
01:26:02.710 --> 01:26:16.319
Redwood C/D - 6504984451: So, yeah, over here, I used UMAP to answer your… did I answer it? Okay, I'm sure you have, just didn't understand the method. Have you only used, optical photometry in this? Okay.

303
01:26:16.420 --> 01:26:19.150
Redwood C/D - 6504984451: Okay, I'll read the paper. Thanks.

304
01:26:25.590 --> 01:26:28.049
Redwood C/D - 6504984451: Just a quick question, why did you…

305
01:26:29.030 --> 01:26:32.189
Redwood C/D - 6504984451: Just a quick question, why did you not use the U-band?

306
01:26:32.870 --> 01:26:48.039
Redwood C/D - 6504984451: Oh, so this is Legacy Survey. Oh, okay, okay, sorry, I thought it's LSSCDP1. Yeah, this is… this is Legacy Survey, but I mean, these are mostly star-forming galaxies, and I mean, if we use, like, the U-band in, like, the DP2, I think that will… that will provide us more information.

307
01:26:53.430 --> 01:26:56.840
Redwood C/D - 6504984451: It sounds like it would be a good… Hackathon.

308
01:26:56.950 --> 01:26:59.569
Redwood C/D - 6504984451: project. Yes. Yeah.

309
01:27:01.320 --> 01:27:19.659
Redwood C/D - 6504984451: relate them directly to each other. Yes, yeah, and also, yeah, and also these depend on, like, noise and resolution, very much so, so, like, comparing with how this, like, depends on, like, noise and resolution across legacy survey and

310
01:27:19.770 --> 01:27:22.909
Redwood C/D - 6504984451: TP2, that would also be something interesting.

311
01:27:26.370 --> 01:27:28.989
Redwood C/D - 6504984451: Okay, well, thank you.

312
01:27:34.450 --> 01:27:44.470
Redwood C/D - 6504984451: Yeah, so, I have just one slide. There's… yeah, that's fine. So, I wanted to discuss the problem, so I'll not sell any solution.

313
01:27:45.260 --> 01:27:49.609
Redwood C/D - 6504984451: So I touched upon this problem yesterday in the strong Lanza session here.

314
01:27:50.960 --> 01:28:01.480
Redwood C/D - 6504984451: And, yeah, so let me start. Like, the name says that there's an image time series, there will be multiple bands, so all of these data points are basically an image in multiple bands.

315
01:28:01.690 --> 01:28:06.400
Redwood C/D - 6504984451: So you know that AISST, like, scans the whole sky in multiple bands. The problem is that?

316
01:28:06.510 --> 01:28:11.239
Redwood C/D - 6504984451: At a given point in time, you have observation in only one band. That makes the…

317
01:28:11.570 --> 01:28:15.320
Redwood C/D - 6504984451: Time series in different bands asynchronously.

318
01:28:17.240 --> 01:28:22.189
Redwood C/D - 6504984451: Now, So, we deal with this for the strong lens transient surge.

319
01:28:22.300 --> 01:28:30.839
Redwood C/D - 6504984451: But I thought that this might be one of the generic problems that the whole LSST transient community should encounter.

320
01:28:31.610 --> 01:28:44.560
Redwood C/D - 6504984451: forget about any logistic challenges, because, like, suppose there are many groups who analyze one frame, maybe a co-add, so this is basically the static, you find VITs, resonates, units, for not.

321
01:28:45.510 --> 01:28:56.669
Redwood C/D - 6504984451: The transient community actually uses this, but you replace each… this n by n pixel image by one photometric data, so that becomes your light curve.

322
01:28:57.960 --> 01:29:01.469
Redwood C/D - 6504984451: But what if you want to analyze the full information?

323
01:29:03.830 --> 01:29:09.750
Redwood C/D - 6504984451: I was searching… The solution for this problem for years.

324
01:29:10.180 --> 01:29:16.130
Redwood C/D - 6504984451: I thought that somebody in the computer fission community, they must have, given

325
01:29:16.650 --> 01:29:23.470
Redwood C/D - 6504984451: Some well-known, well-established foundational model, like some model that we can reuse. We do not need to reinvent the solution.

326
01:29:24.360 --> 01:29:25.849
Redwood C/D - 6504984451: I could not find one.

327
01:29:25.950 --> 01:29:39.639
Redwood C/D - 6504984451: I asked many people, like, in Germany, US, you know, do you know? I could not find it. Okay, so we have solved it, I solved 50%, Claude's solved maybe the other 50%?

328
01:29:39.670 --> 01:29:51.860
Redwood C/D - 6504984451: But the… so my question here is not the solution. My question is that how important this particular question… so I'll ask you a very simple question. How many of you think that these problems should be solved, and could be useful for your science case?

329
01:29:52.790 --> 01:30:00.750
Redwood C/D - 6504984451: Like, just, just, just to clarify, you mean that the question being that, The probabilities for…

330
01:30:00.960 --> 01:30:10.270
Redwood C/D - 6504984451: each time step don't know about. Yes, I'm so sorry I didn't explain the last part. With online prediction, that means…

331
01:30:10.460 --> 01:30:15.290
Redwood C/D - 6504984451: You do not wait for the series to complete, so after each observation, you predict.

332
01:30:15.470 --> 01:30:21.729
Redwood C/D - 6504984451: Suppose you are here, then you are seeing this, and the past is accessible to you.

333
01:30:21.900 --> 01:30:34.919
Redwood C/D - 6504984451: So, you have only data or observation till this point, and based on that, you have to predict. So, naturally, as the observation goes on, goes on, you are accruing more and more data, so your prediction improves and improves.

334
01:30:34.960 --> 01:30:44.569
Redwood C/D - 6504984451: So that, that is the last one. I'm so sorry that I forwarded the, explaining. So, yeah, so anyone thinks that this is worth solving? Yes? I mean, I think in…

335
01:30:47.350 --> 01:31:04.800
Redwood C/D - 6504984451: Like, not in physical systems so much, but in engineering systems, this is a problem that exists and is worked on, where you have multiple sensors that predict… So, all the papers I have… I went through CVPR myself, Claude went through maybe the others.

336
01:31:04.810 --> 01:31:10.900
Redwood C/D - 6504984451: CDPR is their, like, their flagship conference for the computer vision community, and they write

337
01:31:10.950 --> 01:31:20.000
Redwood C/D - 6504984451: Yeah, like, burning questions, the answers there. I couldn't find it. So, if you remove one problem from this particular problem, things become much easier.

338
01:31:20.090 --> 01:31:26.590
Redwood C/D - 6504984451: If you, like, reduce this, From a 2D image to one number, solutions there.

339
01:31:26.720 --> 01:31:38.050
Redwood C/D - 6504984451: If you align the frames, this is the classic video transfer, like the video transfer, and all the self-driving car with different sensors, so the problem is that their frames are aligned, or…

340
01:31:38.430 --> 01:31:50.859
Redwood C/D - 6504984451: they are so, finely spaced that you are allowed to do interpolation. So here, another thing I, forgot to mention, you're not allowed to do any type of interpolition. So you have to deal with only the observation.

341
01:31:52.950 --> 01:32:03.829
Redwood C/D - 6504984451: I think the bottleneck is the image-based problem. I think the bottleneck would be the fact that this is an image-based problem, but there's multi-dimensional

342
01:32:04.080 --> 01:32:05.559
Redwood C/D - 6504984451: Sensors that are not…

343
01:32:05.840 --> 01:32:20.010
Redwood C/D - 6504984451: multidimensional inputs that are not images for which this problem is worked on, right? For the self-driving car, suppose you have multiple cameras, so then that is an image-based problem. But in that case, the frames are…

344
01:32:20.040 --> 01:32:28.940
Redwood C/D - 6504984451: Allied. But suppose you are taking some radar data, which is not… I do not know whether the radars are, like, can be one-dimensional, but of course, there could be some…

345
01:32:28.970 --> 01:32:41.739
Redwood C/D - 6504984451: asynchronous one-dimensional data, and for… for that, the solution exists. The solutions are not groundbreaking. So you just basically need, transformers, you mix the spatial and temporal transformer in some certain

346
01:32:42.090 --> 01:32:56.160
Redwood C/D - 6504984451: Okay? But what surprised me, that I could not find a solution, and yes, so there, like, still, I'm pretty sure there exists some obscure paper, which I could not find. But, yeah, so that actually surprised me.

347
01:32:56.560 --> 01:33:08.170
Redwood C/D - 6504984451: And there was a second slide that's the solution, but yeah, so this is not so, interesting. So you can think of the solution. If you are a master of transformers, then I'm pretty sure, like, if you can…

348
01:33:09.460 --> 01:33:16.989
Redwood C/D - 6504984451: like, design your own attention mechanism, then you can solve it. So there is no, not some issues, and anyway, we are writing paper.

349
01:33:17.560 --> 01:33:23.539
Redwood C/D - 6504984451: That's not a thing. But I wanted to highlight that this problem exists, and if I understand it correctly.

350
01:33:23.720 --> 01:33:28.470
Redwood C/D - 6504984451: Then, that gives you the best solution, like, be it a…

351
01:33:28.810 --> 01:33:34.400
Redwood C/D - 6504984451: Binary classification, be it some sort of regression, be it generation, it doesn't matter.

352
01:33:34.700 --> 01:33:40.519
Redwood C/D - 6504984451: So, how you, like, analyze it, so, yeah. So, I want to edit, sorry.

353
01:33:41.240 --> 01:33:42.780
Redwood C/D - 6504984451: Alex, can I?

354
01:33:43.820 --> 01:33:47.419
Redwood C/D - 6504984451: Yeah, I, I wonder if I… it's,

355
01:33:47.510 --> 01:34:05.350
Redwood C/D - 6504984451: Why does it have to be images? Because… because there are solutions when it's just, like, if it were photometric light curves instead of looking at the images. What do you… do you lose something by using those solutions instead of the images? It depends on the problems, right? So, for example,

356
01:34:06.600 --> 01:34:07.430
Redwood C/D - 6504984451: Thanks.

357
01:34:07.700 --> 01:34:16.529
Redwood C/D - 6504984451: So, we stumbled upon this problem when we were analyzing, we were trying to detect the lens transients as early as possible.

358
01:34:17.120 --> 01:34:33.029
Redwood C/D - 6504984451: But there could be many other problems. So, whenever you reduce the whole information into one number, you must be losing something. So that might not be so important for some science cases. For example, if you are classifying transients from the light curve, that's fine.

359
01:34:33.930 --> 01:34:45.829
Redwood C/D - 6504984451: But there could be some, like, there should be, like, you are always losing some information to get to the solution. But you can also do that classification from the images, and another part is that

360
01:34:46.200 --> 01:34:59.159
Redwood C/D - 6504984451: The future time domain surveys, so this will, I think, again, like, what do I know? But still, like, with my small, experience, I think that this would be a very generic problem that you have to deal with.

361
01:34:59.560 --> 01:35:03.289
Redwood C/D - 6504984451: If you embed your images, if you use an embedder, like, because…

362
01:35:03.830 --> 01:35:05.930
Redwood C/D - 6504984451: So we agreed that the problem…

363
01:35:06.140 --> 01:35:08.450
Redwood C/D - 6504984451: That solutions to the problem exist.

364
01:35:08.550 --> 01:35:23.530
Redwood C/D - 6504984451: If the input is not images, but if you can embed your images in an input that is much simpler while not being a single data point. Doesn't that bring you back to existing solution engineering domain?

365
01:35:24.070 --> 01:35:40.899
Redwood C/D - 6504984451: So, we… the solutions we found are basically this. So, you embed the… so basically a mixture of marriage between VITs and temporal transformers, and basically, it follows this technique. So, it basically not… so we do not embed the full image into one token.

366
01:35:41.160 --> 01:35:52.070
Redwood C/D - 6504984451: Rather, you take a VIT approach, that you patchify the images. So, all these things are technical, right? So, if you are interested, you can always talk to me, you can even talk to Claude, Claude can explain it better.

367
01:35:52.320 --> 01:36:05.109
Redwood C/D - 6504984451: Much better than me. But I just wanted to inform people that this problem… for me, like, I was surprised at why nobody, came a solution to this.

368
01:36:12.590 --> 01:36:14.870
Redwood C/D - 6504984451: Yeah, I'm afraid you can just…

369
01:36:18.310 --> 01:36:23.449
Redwood C/D - 6504984451: I'll say that there's… there's been some research… So, I already showed these yesterday, so…

370
01:36:23.580 --> 01:36:35.259
Redwood C/D - 6504984451: Well, I meant, there's been some recent work by an ISSC member, Grant Joe, who's a grad student at CMU, that… it's not quite… so instead of images, it's like…

371
01:36:35.320 --> 01:36:48.460
Redwood C/D - 6504984451: A multiplied flight curve is one of… can be thought of as multi-channel, and also the probability vector from the previous classifications as another set of channels.

372
01:36:48.520 --> 01:37:00.750
Redwood C/D - 6504984451: Because there's… it's over time, and it's multi-class, so there are… there are higher dimensional methods that deal with this, not from the images, but you… even if you turn the images into, like.

373
01:37:01.240 --> 01:37:09.110
Redwood C/D - 6504984451: it doesn't have to be two parameters, but, like, some… some latent space that can still be a fair number of parameters. There are ways of dealing with this.

374
01:37:09.210 --> 01:37:26.030
Redwood C/D - 6504984451: Yeah, of course, like, if there… so, if there existed a solution, I didn't have to find one. So, my complaint is that I could not find one. It is a very recent paper. So, okay, we'll… Okay, so we'll communicate on this. Thank you.

375
01:37:26.140 --> 01:37:31.760
Redwood C/D - 6504984451: That would be me. Do you mind, refreshing the browser?

376
01:37:43.310 --> 01:37:45.420
Redwood C/D - 6504984451: It should be a slide before that.

377
01:37:46.800 --> 01:37:49.100
Redwood C/D - 6504984451: Oh, there it is. Okay, okay. It's there.

378
01:37:50.000 --> 01:37:51.330
Redwood C/D - 6504984451: Okay.

379
01:37:52.320 --> 01:38:08.390
Redwood C/D - 6504984451: So, kind of part of my responsibility, being a South African astronomer here at CAPAC, where I've been visiting, for 9 months now. I'll be here for another 3 months on my Fulbright, 1-year fellowship, is to disseminate as much information as possible about

380
01:38:08.390 --> 01:38:13.170
Redwood C/D - 6504984451: SKA, square kilometer array, and, developments thereof.

381
01:38:13.170 --> 01:38:28.040
Redwood C/D - 6504984451: since the American community has unfortunately somehow been kind of left out of all of that, given that the USA is not a partner country. I'm probably saying some controversial things, so I'll stop.

382
01:38:28.040 --> 01:38:33.769
Redwood C/D - 6504984451: But, thankfully, there are opportunities for partnership, collaborative.

383
01:38:33.770 --> 01:38:51.380
Redwood C/D - 6504984451: partnerships between Rubin LSST and the SKA, and this has been made possible through the in-kind programs that have been developed. There's several of them. In-kind partnerships with Chile, of course, which is hosting the actual LSST CAM and the observatory itself, but also

384
01:38:51.380 --> 01:38:58.100
Redwood C/D - 6504984451: with South Africa through our in-kind partnership that is operated by the South African Astronomical Observatory.

385
01:38:58.100 --> 01:39:12.670
Redwood C/D - 6504984451: Where we currently have 10 PIs selected, so these are all faculty members at various universities and institutes in South Africa, who have their own scientific foci, and also groups of junior associates, as we call them, students.

386
01:39:12.970 --> 01:39:21.200
Redwood C/D - 6504984451: Or postdocs, early career scientists, and the idea is to make sure that data rights access

387
01:39:21.200 --> 01:39:36.559
Redwood C/D - 6504984451: is made available to, South African, researchers through this program. So, given the timeline of the SKA, optimistically, we expect to have a completed telescope by the early to mid-2030s.

388
01:39:36.560 --> 01:39:50.170
Redwood C/D - 6504984451: So, given also the timeline of Rubin LSST and the decadal survey, there'll be an excellent opportunity to combine deep radio continuum observations, such as radio continuum.

389
01:39:50.170 --> 01:40:05.329
Redwood C/D - 6504984451: polarization products, polarized emission, observations, and also H1, spectral line emission, which will be, more sensitive than has ever been, given the scale of the SKA.

390
01:40:05.740 --> 01:40:13.179
Redwood C/D - 6504984451: So that's happening, and it's something, that if American researchers, I, I'd love you to be aware of.

391
01:40:13.180 --> 01:40:31.310
Redwood C/D - 6504984451: So that you can form partnerships with, astronomers in South Africa. So just a snapshot, an example of something that I've already got a student of mine working on, Maria Zoras, who is at UNISA, which is my home institute. She is now combining radio

392
01:40:31.710 --> 01:40:40.050
Redwood C/D - 6504984451: continuum observations that we already have from Meerkat, which is the 64 antenna, array that has already been commissioned now.

393
01:40:40.230 --> 01:40:41.819
Redwood C/D - 6504984451: Five years ago.

394
01:40:41.820 --> 01:41:02.700
Redwood C/D - 6504984451: And combining this with archival, optical, and near-infrared photometry to try to, firstly label the data, according to empirical models, such as the infrared radio correlation, which gives us our QIR, that allows us to determine where our star-forming galaxies and radio-loud AGN

395
01:41:02.700 --> 01:41:17.330
Redwood C/D - 6504984451: And, other as-yet-unknown classifications of radio if it's galaxies. So this… the whole focus here is galaxies, essentially. So she's, using, the empirical data to inform the models.

396
01:41:17.470 --> 01:41:30.890
Redwood C/D - 6504984451: And, effectively apply a SIMCLR procedure to get a learned representation of the data based on labels that, she's assigned to the data, informed by the empirical

397
01:41:31.370 --> 01:41:35.400
Redwood C/D - 6504984451: Work of the… in the archives, in the… in the literature.

398
01:41:35.540 --> 01:41:43.600
Redwood C/D - 6504984451: And projected it into UMAP 2D space, as we saw in Joy's plot. That's a way to visualize clusters of sources.

399
01:41:43.720 --> 01:42:02.090
Redwood C/D - 6504984451: And then used HDBCAN to find groupings of these sources, and she's been able to, within this UMAP space, find where the star-forming galaxies Radio AGN, radio quiet, radio access. So the idea, what she's doing, since it's just an MSC project, is to see if, the empirical

400
01:42:02.090 --> 01:42:06.800
Redwood C/D - 6504984451: relations can be reproduced by these,

401
01:42:07.280 --> 01:42:21.339
Redwood C/D - 6504984451: algorithms and machine learning methods that have been used time and time again, mainly in industry, but now we're applying them to astronomy cases. So, if the question here is just open for discussion, you're welcome to find me.

402
01:42:21.490 --> 01:42:39.360
Redwood C/D - 6504984451: wherever you see me, and, we want to figure out how, DP2 Rubin photometry will be able to… to possibly improve our ability to find out where all these different classifications are. So that's something that,

403
01:42:39.540 --> 01:42:51.000
Redwood C/D - 6504984451: I will be definitely able to get as much, information on improving from those who've worked on these kinds of projects. Yes, thank you very much.

404
01:42:58.010 --> 01:43:00.720
Redwood C/D - 6504984451: Are there… are there specific questions?

405
01:43:01.380 --> 01:43:03.990
Redwood C/D - 6504984451: Nice. Okay.

406
01:43:04.390 --> 01:43:17.819
Redwood C/D - 6504984451: Thank you for sharing this. This is very exciting, Nabilh. So, over here, so what radio data are you using? Yeah, so the radio data we have is from the MITE survey, so it's a…

407
01:43:18.100 --> 01:43:27.220
Redwood C/D - 6504984451: Meerkat Continuum Survey that aims to target deep drilling fields such as Cosmos, XMM, LSS, Elias S1,

408
01:43:27.320 --> 01:43:38.350
Redwood C/D - 6504984451: And that's at, in the L-band, so between 1 to 2 gigahertz. And we've already got all of that imaging completed, and it's, it's, been calibrated.

409
01:43:38.350 --> 01:43:50.659
Redwood C/D - 6504984451: And from that, we're able to run our source finders to… to pick out where we get contiguous, kind of, 5 sigma plus radio emission to identify where our…

410
01:43:50.660 --> 01:44:00.019
Redwood C/D - 6504984451: radio-emitting galaxies most likely are, and we combine that with multi-wavelength archival data. Yeah, so that's the story, how we get the data.

411
01:44:00.200 --> 01:44:07.239
Redwood C/D - 6504984451: And now we're applying these, methods to try to, develop more automated methods for classification.

412
01:44:07.440 --> 01:44:12.339
Redwood C/D - 6504984451: Because all empirical methods can be very slow when we have large volumes of data.

413
01:44:14.910 --> 01:44:32.699
Redwood C/D - 6504984451: Let me see. If there's… can I ask another question, quick question, yeah. So, when we talk about radio data, it's like, intensity, so, I mean, what will SK provide us, like, intensity, polarization, and will you… will that be helpful in this case? Yes.

414
01:44:32.750 --> 01:44:48.450
Redwood C/D - 6504984451: So the SKA, so you can think of the data products where, and they've been calibrated and imaged as, your continuum imaging, which is your intensity map, so that's the eye parameter, the Stokes eye parameter, and then you'll have the QUV,

415
01:44:48.570 --> 01:45:06.589
Redwood C/D - 6504984451: Which provides your polarization products, and then you've also got your H1 21cm spectral line cubes that will also be provided in 4K mode. So that's the three, kind of, categories that we'll get with SKA, yeah. Sure, thank you.

416
01:45:13.840 --> 01:45:23.389
Redwood C/D - 6504984451: Yeah, I'll hold my questions, because they're not really statistical. No, this is very… this is very cool.

417
01:45:23.670 --> 01:45:27.490
Redwood C/D - 6504984451: Super excited about the possibility of having labels…

418
01:45:27.800 --> 01:45:33.020
Redwood C/D - 6504984451: That, that couldn't have been derived from one dataset, but seeing can we

419
01:45:33.460 --> 01:45:40.700
Redwood C/D - 6504984451: predict from one to another, to see if this is happening in the supermit yet.

420
01:45:41.160 --> 01:45:43.200
Redwood C/D - 6504984451: So yeah, cool.

421
01:45:45.910 --> 01:45:46.980
Redwood C/D - 6504984451: Okay, so…

422
01:45:47.190 --> 01:45:55.070
Redwood C/D - 6504984451: Do I just stop and they hear me? Yeah, cool. So, I was asked to maybe present a slide

423
01:45:55.080 --> 01:46:03.650
Redwood C/D - 6504984451: like, earlier today, so apologies for how rough this is. Essentially, the USDF team is trying something different.

424
01:46:03.650 --> 01:46:25.489
Redwood C/D - 6504984451: We are looking to monitor all the infrastructure that is below prompt processing and alert production, so this is, like, the network, the file system, storage, so that we don't run into problems, during observing. And one of the ways we want to do this is assisted by LLM. We already have code that essentially hits,

425
01:46:25.490 --> 01:46:28.690
Redwood C/D - 6504984451: About 150 of these systems.

426
01:46:28.700 --> 01:46:40.359
Redwood C/D - 6504984451: And checks them for, readiness, for, functionality. But, we're in the process of connecting it up to various LLM

427
01:46:40.450 --> 01:46:43.900
Redwood C/D - 6504984451: MCP sort of situations, and also,

428
01:46:44.330 --> 01:46:55.809
Redwood C/D - 6504984451: The idea, sort of, is that we will have pre-observing checks, like a daytime checkout sort of thing for our infrastructure. We'll have during-observing checks, where some, some, a few tests will run on a loop.

429
01:46:55.990 --> 01:47:03.740
Redwood C/D - 6504984451: And somebody could be notified if something goes wrong. And we'll have post-observing checks to make sure that everything's in a good state at the end.

430
01:47:04.890 --> 01:47:14.349
Redwood C/D - 6504984451: So, essentially, this is just starting out. I've essentially just pinged every system I want to test, and for…

431
01:47:14.350 --> 01:47:29.489
Redwood C/D - 6504984451: people, there's a CLI tool that outputs data that looks kind of like this, right? This is from, like, a post-observing analysis of a day where the Ceph cluster, like, went out, and you can see, like, oh, all this stuff is wrong, and

432
01:47:29.630 --> 01:47:35.140
Redwood C/D - 6504984451: all these other systems went fine. But I figured it might be helpful to ask you guys

433
01:47:35.810 --> 01:47:37.949
Redwood C/D - 6504984451: sort of the questions I have, which are, like.

434
01:47:38.120 --> 01:47:49.439
Redwood C/D - 6504984451: what's a helpful way to display information on, like, infrastructure systems, like, during observing, and, like, who's going to be awake, and how do we get that information succinctly to a person?

435
01:47:49.440 --> 01:48:00.420
Redwood C/D - 6504984451: how can we reasonably compare information that's of, like, wildly different formats, right? Like, I know we have Profana dashboards, there's a lot of them, but we're kind of hoping we can get to really drill down into, like.

436
01:48:01.340 --> 01:48:02.720
Redwood C/D - 6504984451: What is wrong?

437
01:48:03.220 --> 01:48:06.159
Redwood C/D - 6504984451: And across the whole system.

438
01:48:06.410 --> 01:48:12.280
Redwood C/D - 6504984451: And there's a whole follow-up check architecture that we want to enable,

439
01:48:12.410 --> 01:48:15.169
Redwood C/D - 6504984451: This sort of, like, yes, here's the response.

440
01:48:15.350 --> 01:48:18.839
Redwood C/D - 6504984451: And then also, like, my…

441
01:48:19.310 --> 01:48:28.080
Redwood C/D - 6504984451: forays thus far into, like, a local LLM for interpreting this data have really led me to realize that,

442
01:48:29.060 --> 01:48:45.790
Redwood C/D - 6504984451: we have a lot of it. And contact management is really hard. You can't just, like, call one check and have it be, like, processed and tell you the root cause was this one, you know? So, like, how would you sign a system, that can manage, like.

443
01:48:46.390 --> 01:48:53.079
Redwood C/D - 6504984451: metadata on 150 systems, like a thousand-line JSON file. So yeah, this is…

444
01:48:53.250 --> 01:49:09.599
Redwood C/D - 6504984451: Also, if you… if you think monitoring infrastructure systems under AP is a cool idea, and you want to work on that, feel free to let me and Adam and Adi know. But yeah, I was asked if I had anything to bring people.

445
01:49:09.870 --> 01:49:11.899
Redwood C/D - 6504984451: This is what I've got, so…

446
01:49:12.180 --> 01:49:13.959
Redwood C/D - 6504984451: I'm excited to see what you think.

447
01:49:14.920 --> 01:49:15.620
Redwood C/D - 6504984451: Yep.

448
01:49:16.080 --> 01:49:28.159
Redwood C/D - 6504984451: Actually, I think that might be… I think you're in the room. Yeah, how many, how much tokens do you expect to be loaded into contact, say, if you have trouble managing?

449
01:49:28.160 --> 01:49:47.290
Redwood C/D - 6504984451: So, the idea here would be to run something on, like, a local node, like a Gemma 4 model. So I don't know how many tokens that actually expense, or if token… I don't think tokens are involved in this, but I may be wrong. Well, how much… how much tokens are you inputting into keeping contacts?

450
01:49:47.460 --> 01:49:50.870
Redwood C/D - 6504984451: I don't have a good answer to that question, okay.

451
01:49:51.080 --> 01:49:55.770
Redwood C/D - 6504984451: But I will happily talk with you later about it. Sure.

452
01:49:55.950 --> 01:50:12.260
Redwood C/D - 6504984451: because there's, like, ways of people in the industry that manage context with things called, like, Ralph Loops, and compaction, and doing, like, summarize… auto-summarizing of, like, things and context, and then reloading that summary into context, and tricks that people do, so there's, like, a…

453
01:50:12.710 --> 01:50:14.819
Redwood C/D - 6504984451: cool thing about it. Cool, yeah.

454
01:50:14.920 --> 01:50:16.350
Redwood C/D - 6504984451: Yeah, I'll look into it.

455
01:50:16.890 --> 01:50:20.940
Redwood C/D - 6504984451: Yeah, for reference, at this point, I have a JSON file, so… Gotcha.

456
01:50:21.280 --> 01:50:22.650
Redwood C/D - 6504984451: Yeah.

457
01:50:23.260 --> 01:50:24.940
Redwood C/D - 6504984451: Thoughts, questions?

458
01:50:29.350 --> 01:50:41.509
Redwood C/D - 6504984451: I hope on the UX side, for your first bullet, that you actually are talking to the staff who are awake in the middle of the night as to how they want to see the information. Yes.

459
01:50:45.780 --> 01:50:47.459
Redwood C/D - 6504984451: Adam, you have anything to add?

460
01:50:50.030 --> 01:51:05.879
Redwood C/D - 6504984451: Yeah, that's a… it's something that… it's a conversation we've had going for a while now, since even before we started thinking about this in… in the Agentic LLM kind of way, even when we were thinking about just straight-ahead deterministic monitoring and… and,

461
01:51:05.880 --> 01:51:10.050
Redwood C/D - 6504984451: Yeah, I think we realized pretty quickly that anybody who's looking at this as a…

462
01:51:10.070 --> 01:51:26.960
Redwood C/D - 6504984451: They're sitting in a different role, and they're looking at other things simultaneously, and so, yeah, just in terms of the basic monitoring dashboards, we already… those break out into things that you might imagine exposing to the observers at night, which have to be very unobtrusive and only really show them something if they can

463
01:51:27.160 --> 01:51:35.220
Redwood C/D - 6504984451: If it's really something they need to act on, and they know what action they need to take by seeing it, but they definitely don't need to be overwhelmed with a bunch of

464
01:51:35.250 --> 01:51:51.760
Redwood C/D - 6504984451: down in the weeds information about what's going on under the hood. And then there's, like, another level of, sort of, Tier 2 support, maybe, where, okay, there is a problem, and can you at least get a system to triage it to the level of, like, do you need to wake up a network person, or…

465
01:51:51.760 --> 01:52:07.100
Redwood C/D - 6504984451: a telescope person, or what kind of… what subsystem is the problem coming from? And then, yeah, even beyond that, I think it's interesting to see, and this is where we, you know, potentially can go deeper on it at the infrastructure layer within the USDF here, is, like.

466
01:52:07.300 --> 01:52:17.870
Redwood C/D - 6504984451: can you actually have an autonomous expert system go really deep into the, like, diagnosing the root cause in this fairly complex, environment here? So…

467
01:52:18.000 --> 01:52:26.449
Redwood C/D - 6504984451: I think we're enthusiastic, but it's still, still early, and of course, we're still early in operations as well.

468
01:52:27.180 --> 01:52:33.370
Redwood C/D - 6504984451: Yeah, so, what type of safety guardrail do you have, like, and how confident you are?

469
01:52:36.830 --> 01:52:55.290
Redwood C/D - 6504984451: Yeah. Okay. So, I guess I should clarify, I'm asking at a point where I have not connected up an LLM yet. But, this is partly because I've been spending a lot of time on making sure that there are guardrails in place. So, the architecture is,

470
01:52:55.410 --> 01:53:05.799
Redwood C/D - 6504984451: There's, I have a Python tool that I call through CLI that, does all these checks, and then, there's a specific endpoint that,

471
01:53:05.890 --> 01:53:17.560
Redwood C/D - 6504984451: the AIC people use, and those MC… those, are tracked in a session. So I have an orchestrator, which at the… that, like, the system.

472
01:53:17.640 --> 01:53:30.850
Redwood C/D - 6504984451: spun up, it starts a session, and then during that session, there's essentially a hard limit on how many times any of these components can be hit or queried. And there's,

473
01:53:31.040 --> 01:53:39.750
Redwood C/D - 6504984451: like, individually, but also overall, also in a certain time frame. And then, for testing, but also for human diagnosis, I have,

474
01:53:40.130 --> 01:53:43.790
Redwood C/D - 6504984451: I have an interactive mode that I'm using to sort of,

475
01:53:43.990 --> 01:53:58.869
Redwood C/D - 6504984451: put the guard barrels in place, one by one, test them myself, and, like, see that the code path actually works just fine. So, it's all in progress, but I've spent about… Yeah, actually, my fear is that what even beaches the guardrails.

476
01:54:00.190 --> 01:54:08.509
Redwood C/D - 6504984451: I do not know the answer, but I'm also figuring out that how to put, like, a hard guardrail that can never be breached.

477
01:54:09.150 --> 01:54:14.530
Redwood C/D - 6504984451: I'm pretty sure that community knows this, I…

478
01:54:15.750 --> 01:54:28.320
Redwood C/D - 6504984451: So, for example, RF minus… so, RF minus RF dot can never be executed, or whatever, like, the world is destroyed, I don't care. But this… so, one is that you…

479
01:54:28.900 --> 01:54:31.969
Redwood C/D - 6504984451: Alias it, like, so you hired this comment.

480
01:54:32.550 --> 01:54:38.890
Redwood C/D - 6504984451: But these are, like, very elementary stuff. I'm pretty sure that the industry has figured out something better, I guess.

481
01:54:40.100 --> 01:54:58.030
Redwood C/D - 6504984451: Yeah, I guess part of the answer is just to proceed cautiously and start with something that's only offering reports and summaries and perhaps debugging things, and then the next stage might be it's suggesting actions that humans might take, but not actually taking those actions autonomously, and then

482
01:54:58.070 --> 01:55:17.540
Redwood C/D - 6504984451: once we… there's… I guess there's a trust-building exercise, almost, that it's just like you do with people, although it's not exactly people. And yeah, at some point, if it's actually taking autonomous actions, right, then it still needs to be very well prescribed as to what those actions can be, and

483
01:55:17.540 --> 01:55:25.630
Redwood C/D - 6504984451: how its permissions are established and inherited, and how it, you know, making sure it doesn't circumvent them. Yeah, it's… I mean, it seems like a…

484
01:55:25.840 --> 01:55:33.900
Redwood C/D - 6504984451: I don't even… Want to speculate what our cybersecurity colleagues are thinking about all of this right now.

485
01:55:34.140 --> 01:55:36.089
Redwood C/D - 6504984451: Yeah, for now, the flint is…

486
01:55:36.630 --> 01:55:41.179
Redwood C/D - 6504984451: check on things, and get a report. So it's just a lot of queries.

487
01:55:45.550 --> 01:55:49.719
Redwood C/D - 6504984451: I have one question, it's a little bit logistical.

488
01:55:49.960 --> 01:56:07.589
Redwood C/D - 6504984451: So, this… this information, is it possible to make it public and make, like, a little data challenge of, like, come up with a way to visualize and predict on this? Or is it too… too secret to really see?

489
01:56:08.310 --> 01:56:09.160
Redwood C/D - 6504984451: I think.

490
01:56:10.390 --> 01:56:15.929
Redwood C/D - 6504984451: I would be surprised if stuff that's happening in SDF, which is most of this.

491
01:56:16.780 --> 01:56:25.070
Redwood C/D - 6504984451: was publicly available in any way. Even as Ruben and SLAC staff, I had to get, like, a special token to look at Girvana.

492
01:56:25.880 --> 01:56:26.670
Redwood C/D - 6504984451: So…

493
01:56:31.960 --> 01:56:44.080
Redwood C/D - 6504984451: One thing we've talked about doing is going back and assembling the logs and telemetry from past incidents where the root cause is known, and using that, potentially dividing that into training and validation.

494
01:56:44.200 --> 01:56:49.710
Redwood C/D - 6504984451: Samples and seeing what, you know, which types of agents are best at

495
01:56:50.040 --> 01:57:00.249
Redwood C/D - 6504984451: Synthesizing all of that unstructured data and getting to the right answer when we know, based on the history of what happened, what the actual problem was.

496
01:57:00.440 --> 01:57:06.110
Redwood C/D - 6504984451: But, yeah, that's just an idea right now. We haven't… I don't think we've tried to actually assemble that data set.

497
01:57:06.220 --> 01:57:12.799
Redwood C/D - 6504984451: Yeah, we've, we've identified a few, incidents and fixes.

498
01:57:13.220 --> 01:57:18.040
Redwood C/D - 6504984451: Great, you have actually done… that's true, you have done some of this, but I think we haven't…

499
01:57:18.350 --> 01:57:23.869
Redwood C/D - 6504984451: Where are… would we potentially… we'll have to think about this some more… whether… whether we might even be able to…

500
01:57:24.210 --> 01:57:29.229
Redwood C/D - 6504984451: package this and distribute it to other… that's an interesting idea.

501
01:57:29.410 --> 01:57:34.679
Redwood C/D - 6504984451: And then the next question after that is, okay, if you did have something that performs particularly well, then…

502
01:57:35.070 --> 01:57:40.070
Redwood C/D - 6504984451: Like, what's the process of bringing that back inside the fence and integrating it?

503
01:57:41.820 --> 01:57:45.130
Redwood C/D - 6504984451: Looking forward to iterating on this, in the future.

504
01:57:47.450 --> 01:57:48.170
Redwood C/D - 6504984451: Thank you.

505
01:57:48.170 --> 01:57:54.140
Wilson, Tom J.: Alex, can I, interrupt? I think you've missed at least me. This is the last slide on the thing, right?

506
01:57:54.710 --> 01:57:56.210
Redwood C/D - 6504984451: Council?

507
01:57:56.620 --> 01:57:58.810
Redwood C/D - 6504984451: in an unexpected…

508
01:57:58.810 --> 01:58:02.070
Wilson, Tom J.: I got caught in your cached Google Slides.

509
01:58:02.190 --> 01:58:07.769
Wilson, Tom J.: I don't know if you've missed anybody else. I was… I put myself somewhere in the middle, was the mistake I made.

510
01:58:08.160 --> 01:58:10.470
Wilson, Tom J.: Above the red.

511
01:58:11.510 --> 01:58:13.009
Redwood C/D - 6504984451: Have you moved them to the end?

512
01:58:13.500 --> 01:58:14.760
Wilson, Tom J.: No, I…

513
01:58:15.230 --> 01:58:18.459
Redwood C/D - 6504984451: Just quickly move them to the end, I can refresh and…

514
01:58:18.640 --> 01:58:23.219
Redwood C/D - 6504984451: Because my mouse is so tiny on this very.

515
01:58:23.220 --> 01:58:25.730
Wilson, Tom J.: Right. Okay, yeah, I can do it then.

516
01:58:26.140 --> 01:58:26.910
Redwood C/D - 6504984451: Okay, cool.

517
01:58:29.060 --> 01:58:29.700
Wilson, Tom J.: I don't know.

518
01:58:39.530 --> 01:58:40.240
Redwood C/D - 6504984451: Oops.

519
01:58:40.410 --> 01:58:43.289
Wilson, Tom J.: Okay, there we go. I'm at the bottom now.

520
01:58:44.320 --> 01:58:45.180
Redwood C/D - 6504984451: Okay.

521
01:58:45.430 --> 01:58:48.209
Wilson, Tom J.: Hopefully, it was just me that That's why I thought.

522
01:58:54.070 --> 01:58:58.160
Wilson, Tom J.: You can go to next slide, please, as I need the one. I just put that one in to not get caught out.

523
01:59:00.260 --> 01:59:15.800
Wilson, Tom J.: Yeah, hi everyone. I'm sort of coming at this from an ISSC point of view from the other direction, because I've done the ISSC bit, almost, with this work. Apologies to anyone who was in the previous session, where you'll have heard this talk just a second ago.

524
01:59:15.800 --> 01:59:24.410
Wilson, Tom J.: But the very quick summary, now that we're approaching the hour, is that I have…

525
01:59:24.730 --> 01:59:35.859
Wilson, Tom J.: I have done all of the complicated statistical analysis of subtracting out all of the contributions to measured Rubin astrometry that you don't want.

526
01:59:36.030 --> 01:59:52.510
Wilson, Tom J.: And I'm now hopefully comparing, as well as you can, the centroid precisions… the centroids measured by the Rubin pipeline to the formal covariance matrices, as also determined by the pipeline.

527
01:59:53.070 --> 02:00:03.369
Wilson, Tom J.: Where essentially you… you would envision some sort of Monte Carlo simulation of repeat measurements which build you a standard deviation of… of some sort of truth error residual.

528
02:00:03.480 --> 02:00:12.400
Wilson, Tom J.: And you compare that to the covariances, and you hope that Y equals X, and therefore your scatter in your positions equals your uncertainty.

529
02:00:12.620 --> 02:00:28.940
Wilson, Tom J.: And the very high-level summary here is that it doesn't… that we fit this very simple model in the middle of the slide there, adding in quadrature some bright and systematic with a slope-corrected covariance between our

530
02:00:29.570 --> 02:00:42.070
Wilson, Tom J.: these empirical scatter and error residuals, and the quoted pipeline uncertainties. And in very small figures, because I had to put all of them on this one slide, you see that in one side we have

531
02:00:42.420 --> 02:01:00.239
Wilson, Tom J.: M of 1, if we use the faint data from Hubble, so we're down at 26th, 27th magnitude. So M is 1, which is great. You can ignore the N on the right-hand side, but when we, for the brighter sources, brighter than 20th magnitude in, Gaia and LSST,

532
02:01:00.240 --> 02:01:11.310
Wilson, Tom J.: We see a, a systematic, slope in the faint but 20th magnitude, data, and a, a bright end,

533
02:01:11.340 --> 02:01:17.310
Wilson, Tom J.: flaw to the astronometry of something like 5 or 10 milliarseconds.

534
02:01:17.490 --> 02:01:34.910
Wilson, Tom J.: But of course, the problem I have is that I've just done an end-to-end test, so I can't answer why this is true. And I sort of wanted to come from this, the other side, and ask, is anyone else looking into the LSST astrometry, for any of the data previews, one or, you know, two, two going forwards?

535
02:01:34.940 --> 02:01:38.790
Wilson, Tom J.: And, you know, maybe we can sort of meet in the middle, where I've got

536
02:01:38.940 --> 02:01:48.329
Wilson, Tom J.: I've got an answer, but I don't know why, and if there are other people who are looking at astronometry from a why-are-things-happening point of view, we had,

537
02:01:48.460 --> 02:02:02.410
Wilson, Tom J.: I talk in the SMWLV session, from Clare Saunders about, the turbulence model, for example, the Gaussian processing fit that should subtract out the… these residuals, and, you know, they will make a difference. So things like that.

538
02:02:02.510 --> 02:02:06.769
Wilson, Tom J.: It was more just a plug for… for friends, more than anything.

539
02:02:07.090 --> 02:02:12.670
Wilson, Tom J.: If anyone was, interested in astronometry from an ISSE point of view.

540
02:02:17.810 --> 02:02:24.360
Wilson, Tom J.: Or are you all galaxy people and don't care where your things are, so long as they're uniformly distributed in the sky?

541
02:02:29.390 --> 02:02:32.360
Redwood C/D - 6504984451: Okay, thank you, this is really interesting. I…

542
02:02:32.580 --> 02:02:39.469
Redwood C/D - 6504984451: People just be impressed that this session has covered so many different areas, like, there's… there's…

543
02:02:39.740 --> 02:02:45.019
Redwood C/D - 6504984451: The different areas of science we can do, also how the data is processed, the intermediate

544
02:02:45.130 --> 02:02:47.110
Redwood C/D - 6504984451: But fundamental data products.

545
02:02:48.720 --> 02:02:53.400
Redwood C/D - 6504984451: images versus… Things that you can put in catalogs.

546
02:02:53.620 --> 02:02:58.320
Redwood C/D - 6504984451: All… all of it has come up, so, we're out of time.

547
02:02:59.420 --> 02:03:00.390
Wilson, Tom J.: So, if you don'

548
02:03:00.770 --> 02:03:05.049
Wilson, Tom J.: at the bottom, you'll find my email address, you can email me if you… if you want to be friends.

549
02:03:06.430 --> 02:03:07.529
Redwood C/D - 6504984451: Also, yeah…

550
02:03:07.660 --> 02:03:14.630
Redwood C/D - 6504984451: through the ISSC, all things are possible. You can follow up on Slack, we can… we can keep the discussion going.

551
02:03:14.990 --> 02:03:33.979
Redwood C/D - 6504984451: I just wanted to bring up a quick point with Sara Bonito, who wanted to be here but could not be because timings, she's joining virtually. But she faced one problem where she found some NEN measurements in RubenLSST data, specifically when looking at the magnitudes for Carina Nebula.

552
02:03:34.040 --> 02:03:46.589
Redwood C/D - 6504984451: And she just wanted to raise that, and maybe, I don't know if others have seen that problems or experiencing that, but if you have, if that's something on your mind, I think that's also something to think about.

553
02:03:46.690 --> 02:03:58.880
Redwood C/D - 6504984451: Okay, so Sarah's not here. Yeah, she's not here, she just asked me to sort of raise this. If you can, forward it along to ISSC, ask the ISSC, if that's a good place to continue the discussion.

554
02:04:00.550 --> 02:04:01.280
Redwood C/D - 6504984451: Cool.

555
02:04:02.460 --> 02:04:09.690
Redwood C/D - 6504984451: Thank you, everyone, but especially Sid and Tom for filling in as, Co-facilitators.

556
02:04:19.600 --> 02:04:20.789
Redwood C/D - 6504984451: Governor's face.