WEBVTT

1
00:00:02.160 --> 00:00:18.339
STFC-RAL-CR03  R61: Welcome to today's seminar. Welcome to Sarah from SEL. So, let me start with a short introduction. Sarah is a faculty. She received her master's degree from University of Oxford.

2
00:00:18.480 --> 00:00:21.000
STFC-RAL-CR03  R61: And a PhD from UCL.

3
00:00:21.100 --> 00:00:31.459
STFC-RAL-CR03  R61: She then joined as a postdoc at Ross Seller University and UCL, before being awarded a Royal Society University Research Fellowship.

4
00:00:31.470 --> 00:00:43.019
STFC-RAL-CR03  R61: Currently, she is a Senior Research Fellow at UCL, and co-director of UCL's Phantom Computing and Phantom Communications Center for Doctoral Training.

5
00:00:43.170 --> 00:00:50.069
STFC-RAL-CR03  R61: She, explores the interface between quantum computing and quantum physics.

6
00:00:50.680 --> 00:00:52.630
STFC-RAL-CR03  R61: Indeed.

7
00:00:52.780 --> 00:00:54.510
STFC-RAL-CR03  R61: the financial techniques.

8
00:00:54.760 --> 00:01:00.590
STFC-RAL-CR03  R61: or accelerate with Discovery Physics beyond this.

9
00:01:00.870 --> 00:01:02.970
STFC-RAL-CR03  R61: Welcome.

10
00:01:04.360 --> 00:01:08.209
STFC-RAL-CR03  R61: Great, thank you so much for the invitation. I'm happy to be there.

11
00:01:08.250 --> 00:01:26.160
STFC-RAL-CR03  R61: Yeah, thanks for the introduction. So, I did actually… I've been in the field for about two decades now, and I spent some time at CES. I did my PhD on the Tevatron experiment, so I've got, you know, Sam here from those days. I know people from, from Tevatron, and then I spent some time at CMS.

12
00:01:26.160 --> 00:01:30.840
STFC-RAL-CR03  R61: Before becoming interested in all kinds of technology in general, for instance.

13
00:01:30.840 --> 00:01:41.430
STFC-RAL-CR03  R61: about 4 years ago. So I will go into Phantom Center, give us a little piece of tabletop acronym to see if we put the dark matter of dark energy via those.

14
00:01:41.430 --> 00:01:51.750
STFC-RAL-CR03  R61: And then, because there wasn't anybody looking into climate computing, at least in the biophobics domain in the UK, I started a group on that. So you'll see

15
00:01:51.750 --> 00:02:04.529
STFC-RAL-CR03  R61: I'm going to sort of give you an overview of what's going on, and mostly in quantum computing and particle physics, and not just with the, you know, results from my research, but also just overall.

16
00:02:06.130 --> 00:02:16.180
STFC-RAL-CR03  R61: And if I understand correctly, there aren't many people in the room who are actively working in quantity computing, or have expertise in that room. So I'll give a brief introduction to…

17
00:02:16.290 --> 00:02:17.439
STFC-RAL-CR03  R61: Do that as well.

18
00:02:17.510 --> 00:02:33.310
STFC-RAL-CR03  R61: So before we get into quantum computing, I mean, so they say that, you know, currently quantum computers are at probably a similar stage as classical computers were back in the 1950s. So if we look at the trajectory of, classical computers, so

19
00:02:33.310 --> 00:02:48.589
STFC-RAL-CR03  R61: In the 1950s, you sort of needed a big room to house them, and apparently the head of IBM or something at the time said that you don't need… he doesn't see a market for more than 5 of these, and you certainly don't need one in your home, etc. And so, we have been very, very…

20
00:02:48.660 --> 00:03:06.799
STFC-RAL-CR03  R61: it has been very challenging for us to predict what the future of these technology is going to be like, and so it's inherently difficult for us to, you know, map out the trajectory of quite everything and say, well, you know, this is definitely going to be the most transformative technology of the 21st century. We don't yet know, but because we… some of the most

21
00:03:06.880 --> 00:03:10.580
STFC-RAL-CR03  R61: transformative technologies that are kind of find a lot of age, like…

22
00:03:11.260 --> 00:03:23.089
STFC-RAL-CR03  R61: classical computing, even AI, the World Wide Web, etc, we couldn't really have both students find that they were going to have such an impact. And so, my…

23
00:03:23.920 --> 00:03:31.499
STFC-RAL-CR03  R61: thing I've been trying to tell people, especially in the UK, is that if you do want to get in on a technology, the best time to do so is when it's in that.

24
00:03:32.310 --> 00:03:49.310
STFC-RAL-CR03  R61: When I say that slow growth phase at the beginning, before it gets into that exponential progress, before it, you know, plateaus, and we're seeing that now with the classical computers, because Moore's Law is slowing down, the systems are becoming so small that we're beginning to experience quantum effects.

25
00:03:49.610 --> 00:03:55.120
STFC-RAL-CR03  R61: So quantum computing right now is sort of in that slow growth phase, and…

26
00:03:55.670 --> 00:03:59.979
STFC-RAL-CR03  R61: Would be a good time if you want to establish leadership and then to get it.

27
00:04:03.570 --> 00:04:15.530
STFC-RAL-CR03  R61: So, the key difference between the two types of competing paradigms is that in the classical case, you have to work with the classical lick, which is your fundamental unit of information.

28
00:04:15.560 --> 00:04:25.909
STFC-RAL-CR03  R61: And the classical bit can only take on the values 0 and 1, whereas in the quantum case, you have a qubit which can take on not just the values 0 and 1, but also a equal position of those values.

29
00:04:25.910 --> 00:04:38.769
STFC-RAL-CR03  R61: So anything, any state that resides on the surface of that log sphere is accessible to you while you're making your computation, while at the end, everything has to collapse, so you will get one. It's just a zero one. So you have…

30
00:04:38.770 --> 00:04:44.049
STFC-RAL-CR03  R61: Access to the entire workspace while you're… while you're doing your computation.

31
00:04:44.060 --> 00:04:54.090
STFC-RAL-CR03  R61: So for two cubis system, you have access to four raker states for… and then it scales up. For an N-cubris system, you have access to those two to the n-dimensional hookers space.

32
00:04:54.090 --> 00:05:11.590
STFC-RAL-CR03  R61: And so the power of quantum computing does come from this exponential increase in the size of the participant, how available to you. But, you know, it's up to you how you sculpt that system to be able to give you an output that's going to give you an advantage over that scope, because at the end, like I said, when you make a measurement, you still only get a zero from that.

33
00:05:14.910 --> 00:05:31.339
STFC-RAL-CR03  R61: So there's two different, two main classes or paradigms of, platinum computing right now. So one is called the platinum leaders. So in this case, these are usually characterized by a large number of noisy qubits, so you have heard of these machines, you've heard of something like, you know, 5,000 qubit machines.

34
00:05:31.340 --> 00:05:45.559
STFC-RAL-CR03  R61: D-wave is one company that specializes in these kinds of money layers. So they're good for solving three, specific problems. So, over here, what you do is you're looking for… you have a thousand-look keyword system.

35
00:05:45.850 --> 00:05:51.760
STFC-RAL-CR03  R61: you're finding… we're trying to find low-energy configurations of some complex energy Hamiltonian.

36
00:05:51.790 --> 00:06:06.459
STFC-RAL-CR03  R61: And you're doing that by using quantum tunneling to escape any local minima. So, basically, your optimization problem is encoded onto these multiple qubits, and you find the, the sort of ground state of that Hamiltonian.

37
00:06:06.460 --> 00:06:11.830
STFC-RAL-CR03  R61: And, that should be, probably in Microsoft, the optimization.

38
00:06:12.060 --> 00:06:23.160
STFC-RAL-CR03  R61: I'm going to talk mostly today for, everything that I'm going to refer to as going to be for active gate surface, so this is very similar to the way that we do fastball machines.

39
00:06:23.340 --> 00:06:32.639
STFC-RAL-CR03  R61: We have some input qubits, and you make a set of… apply a set of unit-free transformation to those qubits to a series of gates.

40
00:06:32.890 --> 00:06:51.860
STFC-RAL-CR03  R61: And these are typically characterized by a small number of keywords, but they are universal in that you should be able to put any kinds of problems onto that device. And the sort of big companies interested in this are Google, IBM, Microsoft, and also a lot of the smaller startups, like Yeti.

41
00:06:52.700 --> 00:06:55.460
STFC-RAL-CR03  R61: They're doing, like, database practices.

42
00:06:56.580 --> 00:07:03.239
STFC-RAL-CR03  R61: So, like I said, the structure is very similar to the classical one, so you can put a series of, series of,

43
00:07:03.880 --> 00:07:22.560
STFC-RAL-CR03  R61: a series of qubits, and we apply a set of gates, which should be unitary, and then we make a measurement at the end. The difference here is, of course, that you're now leveraging key aspects of the backup phenomena that are not accessible to you in the passcode sheet. So, for instance, your position, and entanglement.

44
00:07:23.350 --> 00:07:28.240
STFC-RAL-CR03  R61: So, some of the most important gates are the ones that allow you to do that.

45
00:07:28.240 --> 00:07:45.750
STFC-RAL-CR03  R61: So this is… the Heimod Gate is one of the most frequently used ones. It's going to be the first one that's used in a financial circuit. It doesn't have a classical equivalent, and what it does is that it initializes… it puts a qubit that's initialized in either the 0 or 1 state into a superposition of constitutes.

46
00:07:47.090 --> 00:07:52.390
STFC-RAL-CR03  R61: Then we have a C-note gate, and these are just some of the, walking gates.

47
00:07:53.430 --> 00:08:04.520
STFC-RAL-CR03  R61: So we have seen on gate, and this is a two-qubit operation, so it flips a thing, a target qubit brings to that. A controlled one. This is used to create, tangled gains, entangled qubits.

48
00:08:04.710 --> 00:08:14.130
STFC-RAL-CR03  R61: And an extension of that is a Polygate, which does the same thing, but, or a 3Q, so it entangles 3, it's capable of entangling.

49
00:08:14.250 --> 00:08:15.480
STFC-RAL-CR03  R61: 3 qubits.

50
00:08:18.050 --> 00:08:34.200
STFC-RAL-CR03  R61: So, there are a lot of different platforms, so as you can… as you've probably heard, there's a lot of companies involved in this. We have different types of QBricks, so there's lots of different, platforms with their own pros and cons, their own engineering challenges.

51
00:08:34.730 --> 00:08:36.750
STFC-RAL-CR03  R61: And there's currently no clear front.

52
00:08:36.900 --> 00:08:46.439
STFC-RAL-CR03  R61: frontrunner. So we don't know which of these is going to become the breakthrough technology that might give us, you know, ball-moderate, universal platinum computers.

53
00:08:46.610 --> 00:08:54.340
STFC-RAL-CR03  R61: Superconducting qubits are being… pioneered by some of the big tech giants, Google, IBM is also Getty.

54
00:08:54.470 --> 00:09:00.300
STFC-RAL-CR03  R61: Oxford Plan Circuit as well, which is a smaller… A smaller startup.

55
00:09:00.410 --> 00:09:14.520
STFC-RAL-CR03  R61: There's trapped ions photonics, which may conceivably give us to be able to do quantum computing at room temperature, so at least that's what they're trying to do. There's a company in Canada, it's an audience that's working on that.

56
00:09:14.550 --> 00:09:26.530
STFC-RAL-CR03  R61: and neutral atoms, and also other ones, too. So, again, there's lots of different platforms for these, and we don't yet know which of these types of qubits is going to become.

57
00:09:26.810 --> 00:09:29.299
STFC-RAL-CR03  R61: It's going to become the wake of technology.

58
00:09:36.380 --> 00:09:55.399
STFC-RAL-CR03  R61: So, we've probably heard… you've probably heard a number of, sort of, planes in the news about active supremacy, you know, have we… have we done something that's gone beyond, what we could do with classical machines? So this was an article, this was from quite some time ago, I think in 20… so 2019, where

59
00:09:55.630 --> 00:10:13.749
STFC-RAL-CR03  R61: Google claim that they've achieved quantum supremacy. So this is the benchmark of doing something on a quantum device that's pretty much intractable in a feasible scale of time in a classical one. So they claim that with their 54-qubit quantum machine, they could do this random sampling calculation in just over 3 minutes.

60
00:10:14.170 --> 00:10:28.149
STFC-RAL-CR03  R61: Something that they would have expected classical machine to take 10,000 years. However, there was a counterclaimed by IBM some time later to say that actually, with some clever crafting, they could do this on a classical machine in two and a half days.

61
00:10:28.280 --> 00:10:41.589
STFC-RAL-CR03  R61: So all this today, I mean, I think there have been several claims given after this. This was relatively old now, which you'll notice in my computing is that, you know, even a seminar that I gave a year ago, I'll have to update it, just because the field isn't moving so fast.

62
00:10:42.360 --> 00:10:46.150
STFC-RAL-CR03  R61: So… Something to note is that

63
00:10:46.890 --> 00:10:56.689
STFC-RAL-CR03  R61: When you're using classical computing as benchmark, you're chasing a moving target, because as your quantum machines are getting better, so are your classical hardware and also the algorithms.

64
00:10:56.840 --> 00:11:06.349
STFC-RAL-CR03  R61: So, I think these were… IBM were able to do this with, just a fancier, more sophisticated algorithm. They were able to say that, actually, instead of 10,000 years, we can do it in 2005 days.

65
00:11:07.560 --> 00:11:17.399
STFC-RAL-CR03  R61: And the second thing to note, this was some hype curve model that they put out for Latin computing, is that there is obviously some hype.

66
00:11:17.450 --> 00:11:33.869
STFC-RAL-CR03  R61: And this is generated by the fact that you have these big tech giants, you know, buying for supremacy, you have a lot of smaller startups buying more VC funding, etc. So there is going to be some degree of hype, which is not altogether bad for the, for getting the technology to advance.

67
00:11:34.320 --> 00:11:45.379
STFC-RAL-CR03  R61: But it does mean that you have heightened expectations, and if they're not met, then we might be… we are not talking about a planting winter, so it's akin to what we saw for AI, where, you know, AI had this…

68
00:11:45.380 --> 00:11:55.949
STFC-RAL-CR03  R61: I don't know, 1960s, or whatever it was, 1980s, you know, AI was expected to do great things, and then it… it did fulfill those expectations, and money pulled out, and then…

69
00:11:55.950 --> 00:12:13.119
STFC-RAL-CR03  R61: It was, you know, people are still working on it, but, at least in the public consciousness, it kind of, dissipated. So people are now worried about planting winter just because of the, sort of, heightened expectation that any day now, planting procedures are going to, break, encryption and,

70
00:12:14.250 --> 00:12:22.940
STFC-RAL-CR03  R61: do all these wonderful things that, that you can't do with classical machines. So, just to say, there is some degree of height, but, that's…

71
00:12:23.240 --> 00:12:25.070
STFC-RAL-CR03  R61: In some ways, to be expected.

72
00:12:25.650 --> 00:12:28.640
STFC-RAL-CR03  R61: In terms of potential applications, there's…

73
00:12:28.960 --> 00:12:47.479
STFC-RAL-CR03  R61: a lot of applications that have been identified. We don't know which of these might be… some of them might be easier and more doable in the near term than others. So, for instance, one of the first use cases is considered to be some chemistry, where you can simulate multi-qubit systems, and that actually gives you a big advantage in being able to,

74
00:12:47.500 --> 00:12:50.090
STFC-RAL-CR03  R61: Accelerate the drug discovery, etc.

75
00:12:50.470 --> 00:12:56.300
STFC-RAL-CR03  R61: Then would something like hundreds to thousands of keywords, you might be able to solve optimization problems.

76
00:12:56.420 --> 00:13:01.259
STFC-RAL-CR03  R61: People had looked at, you know, optimizing transport groups and, so…

77
00:13:01.780 --> 00:13:20.729
STFC-RAL-CR03  R61: optimizing portfolios and financial services, etc. And then, machine learning as well. So a lot of the things that we've been doing in particle physics have been at the intersection between computing and machine learning, putting together all of the buzzwords you can possibly… you can possibly get. But these are considered to be, sort of.

78
00:13:21.330 --> 00:13:24.859
STFC-RAL-CR03  R61: midterm, let's say, application of, of course.

79
00:13:27.360 --> 00:13:32.569
STFC-RAL-CR03  R61: So, in terms of investment, there has been a lot of investment from both governments and,

80
00:13:33.200 --> 00:13:49.420
STFC-RAL-CR03  R61: external, the private sector, so you can see the, sort of, investment chart go up over here, and the UK announced, I think, which budget, it was not so long ago, they announced that they were going to invest 2.5 billion, pounds over the next 10 years or so.

81
00:13:49.910 --> 00:14:03.490
STFC-RAL-CR03  R61: And, this has been, you know, there'll be $200 million in venture capital, funding, and about 40 UK small, medium-sized enterprises are focused on planning. So there is…

82
00:14:03.660 --> 00:14:13.629
STFC-RAL-CR03  R61: There is a lot of emphasis on the, you know, money side of things, and one of the biggest bottlenecks that's been identified is that there's an acute skills shortage. So the number of jobs that are

83
00:14:13.760 --> 00:14:18.749
STFC-RAL-CR03  R61: King of Water Plants and skilled people are out… are actually outweighing the number of people

84
00:14:18.850 --> 00:14:21.490
STFC-RAL-CR03  R61: that are there, fill them by, like, to each one.

85
00:14:21.680 --> 00:14:24.360
STFC-RAL-CR03  R61: Like, 3 to 1 or 2 to 1, sort of skills gap.

86
00:14:24.460 --> 00:14:28.959
STFC-RAL-CR03  R61: So the government has identified this need at the point where we work.

87
00:14:33.100 --> 00:14:45.879
STFC-RAL-CR03  R61: So, what are we doing with this in particle physics? So we have, so right now, it's just, it's in a very exploratory phase, so I'm not going to be able to come and tell you that actually, you know.

88
00:14:45.980 --> 00:14:51.949
STFC-RAL-CR03  R61: We've done something that, that you can do with a classical machine, or that you can actually

89
00:14:52.460 --> 00:14:55.159
STFC-RAL-CR03  R61: Put it right down into your pipeline today.

90
00:14:56.170 --> 00:15:11.569
STFC-RAL-CR03  R61: But it's an exploratory stage because the quantum devices are such, they're small, they're, you know, they're what your… what your classical computers were doing in the 1950s, so I can't compare that to an iPhone right now. It's sort of similar with classical and quantum devices.

91
00:15:11.730 --> 00:15:15.719
STFC-RAL-CR03  R61: So all we're doing right now is trying to figure out what kind of

92
00:15:15.910 --> 00:15:21.260
STFC-RAL-CR03  R61: problems and particle physics might be more… Or…

93
00:15:21.400 --> 00:15:38.040
STFC-RAL-CR03  R61: for quantum algorithms, for being able to just use quantum algorithms to speed them up, or to do things that we couldn't really do with a classical machine. So these are some of the things that my group has looked at, looking at calculating stacking amplitudes, simulating carbon showers in a novel way.

94
00:15:38.100 --> 00:15:43.830
STFC-RAL-CR03  R61: We did some sort of cross-al work with the astrophysics group at UCL for classifying black galaxies.

95
00:15:44.000 --> 00:15:49.840
STFC-RAL-CR03  R61: And, reconstructing particles tracks and also looking at, nuclear surgery.

96
00:15:50.790 --> 00:15:54.290
STFC-RAL-CR03  R61: And… Yeah.

97
00:15:54.860 --> 00:15:55.810
STFC-RAL-CR03  R61: Terrific.

98
00:15:57.000 --> 00:16:16.970
STFC-RAL-CR03  R61: So the first one I'll talk about is track reconstruction. So this is one of the key challenges that the upgrade is that I see is going to be doing track reconstruction, because we have a very dense environment, it's a very busy, high pilot environment, which is over, up to 200 overlapping protocols, so you need a lot more CPU and storage. And so one of the

99
00:16:17.410 --> 00:16:21.250
STFC-RAL-CR03  R61: Obvious questions that are being asked, you don't just help.

100
00:16:23.280 --> 00:16:25.310
STFC-RAL-CR03  R61: So this is, you know, a…

101
00:16:26.060 --> 00:16:37.930
STFC-RAL-CR03  R61: This is a connecting the dots exercise, so you have something like this, you have 10,000 hits in the detector, this is what it would look like, and then your job is to try to connect those histories in the right way to make the particle track.

102
00:16:38.230 --> 00:16:54.150
STFC-RAL-CR03  R61: So this, you know, you have to do… the combinatorics are huge, and the number of the… if you increase the number of hits in your detector, of course, it sort of explodes. So here, you have to go from 100,000 hits in your detector to about 10,000 or so particle tracks.

103
00:16:55.210 --> 00:16:57.940
STFC-RAL-CR03  R61: So that's the… that's the challenge.

104
00:16:58.670 --> 00:17:01.350
STFC-RAL-CR03  R61: So we did this,

105
00:17:01.840 --> 00:17:21.249
STFC-RAL-CR03  R61: we looked at a very small part of this problem, and that's what you have to do with the machines right now. You can't just take the whole problem and then try to solve it. The whole… the whole aim here is to take some pipeline, some sort of components of your pipeline and see if you can do that better than your past machines. And so we reformulated this

106
00:17:21.520 --> 00:17:33.200
STFC-RAL-CR03  R61: traffic instruction as a classification task, and simply asked, does a particular track segment, which in this case consists of three consecutive hits, does it belong to a particle track or not?

107
00:17:33.720 --> 00:17:44.349
STFC-RAL-CR03  R61: And we did this using a machine learning algorithm called a support vector machine. So this is a… an algorithm that tries to find the optimal boundary between classes of data.

108
00:17:44.560 --> 00:17:46.139
STFC-RAL-CR03  R61: And it…

109
00:17:46.370 --> 00:18:02.879
STFC-RAL-CR03  R61: does this by finding this optimal hyperplane that maximizes the margin between the two classes of data, where these are the ones, using only the closest data points here, and we can call this four vectors. So it maximizes the margin and tries to separate the two

110
00:18:03.140 --> 00:18:04.749
STFC-RAL-CR03  R61: To pass the data.

111
00:18:07.050 --> 00:18:13.689
STFC-RAL-CR03  R61: So we used a… a hybrid… Classical Quantum Machine Learning Algorithm.

112
00:18:13.790 --> 00:18:36.019
STFC-RAL-CR03  R61: which looks at… which uses a quantum kernel, so it uses code vector machine with a quantum kernel. So I'll explain what that… what I mean by that. So, basically, you have some input screens for. These are your two paths of data, and the decision boundary separating those two paths of data is very, very complicated. It's not… they're not linearly separated. So what you do is you use a kernel.

113
00:18:36.180 --> 00:18:42.280
STFC-RAL-CR03  R61: essentially gives you a similarity measure between two base points. It tells you how close two base points are.

114
00:18:42.400 --> 00:18:48.620
STFC-RAL-CR03  R61: and you allow the SVM to operate this using a kernel, you're able to go to this from a higher dimensional

115
00:18:48.920 --> 00:18:59.870
STFC-RAL-CR03  R61: Hilbert, some higher dimensional space, where these two classes of data become linearly separable. So you go from some input space to a feature space where your

116
00:19:00.010 --> 00:19:11.349
STFC-RAL-CR03  R61: data are going to draw a line or a boundary as sort of a clean boundary between your two paths of data, whereas they were very complex to separate in your input space.

117
00:19:11.350 --> 00:19:22.720
STFC-RAL-CR03  R61: And we've done this using a quantum kernel, because what the quantum kernel… basically, once you encode your data onto a quantum device, you are implicitly going into some higher-dimensional feature space.

118
00:19:23.690 --> 00:19:28.559
STFC-RAL-CR03  R61: And so you're able to separate your data much cleaner in this, in this hybrid space.

119
00:19:28.890 --> 00:19:41.680
STFC-RAL-CR03  R61: And this was inspired from a paper that showed that actually using such techniques, you can efficiently exploit this higher dimensional space and generate maps that are very difficult for classical

120
00:19:41.730 --> 00:19:52.870
STFC-RAL-CR03  R61: classical algorithms do. And so they used two different types of artificially generated data with differing levels of complexity. And they found that using this technique, the quantum

121
00:19:53.100 --> 00:20:08.479
STFC-RAL-CR03  R61: sort of algorithm was flexible enough to draw decision boundaries between, sort of, simple, simple cases and also much more complicated data sets. So it was much more flexible in being able to draw those, complex and also simple decision boundaries.

122
00:20:10.100 --> 00:20:12.710
STFC-RAL-CR03  R61: And this has been a general takeaway from

123
00:20:12.850 --> 00:20:22.169
STFC-RAL-CR03  R61: This has been a general takeaway for some of the quantum algorithms people have looked at. I… they've asked, can we draw more complex decision boundaries between class and beta?

124
00:20:23.680 --> 00:20:40.169
STFC-RAL-CR03  R61: So I won't go through the full details, the details of this, but we used a TrackML dataset, and we looked at all this, and we just set up two quesector pits. So there was some pre-processing done on that data to try to reduce the bulk of the data.

125
00:20:40.170 --> 00:20:52.040
STFC-RAL-CR03  R61: And, get some triplers, so these are set up 3 hits. And then we put that through a classical algorithm, and then a quantum algorithm, and then looked at the difference between

126
00:20:52.110 --> 00:21:09.699
STFC-RAL-CR03  R61: all three of these. So there's a classical algorithm, part-to-one, and also it's the randomly selected triplets, because we wanted to make sure that this pre-processing step that we've done over here doesn't already give you very, very good triplets. So you want to make sure that you're… you're doing better than just randomly selected triplets from this pre-processed data.

127
00:21:11.260 --> 00:21:28.499
STFC-RAL-CR03  R61: So this is what the quantum circuit looked like for computing that kernel. We're not gonna go into details over here, but it's essentially the spatial heads for all three, all three, spatial coordinates for all three of your heads, so you'd use nine qubits overall to encode the data.

128
00:21:29.350 --> 00:21:46.600
STFC-RAL-CR03  R61: We did this in two different regions. One was the barrel region, and then we wanted to look at the innermost region as well. This is, you know, close to your bean pipe, where it's very dense, it's much more complex, the stuff that's going on over here. And we wanted to see whether the quantum algorithm is able to draw a

129
00:21:46.600 --> 00:21:52.900
STFC-RAL-CR03  R61: Better decision boundary in this much more, dense, much more messy phase space.

130
00:21:53.270 --> 00:22:07.459
STFC-RAL-CR03  R61: So, this is what you can see for the barrel regions. This is the… this is what you would get if you were just randomly picking from your… from your samples. This is the red line over here. So, as you can see, most of the algorithms are doing better than that, so it's not just that our green postage data is already very nicely.

131
00:22:07.910 --> 00:22:27.240
STFC-RAL-CR03  R61: The quantum is the blue, and the yellow is the… is the classical. So they do… they're doing a very, you know, comparable job in the barrel region, but then they're most layers, so this is where you go right down to the… that very messy, very close to the weed pipe region. You can see the quantum might be slightly… there is some slight…

132
00:22:27.810 --> 00:22:33.509
STFC-RAL-CR03  R61: Performance, improvement over the, the FASCO one.

133
00:22:35.900 --> 00:22:41.800
STFC-RAL-CR03  R61: So then we use the same algorithm for something in, we're looking at,

134
00:22:41.860 --> 00:22:45.169
STFC-RAL-CR03  R61: That's fine, Galaxy. I'm not gonna go into

135
00:22:45.210 --> 00:23:02.860
STFC-RAL-CR03  R61: detail in this one, but it was essentially a very similar thing. It's just to show that there is actually a lot of crossovers. Once you develop the techniques in one discipline, you can port them over to another one, and this is similar to what's going on with AI as well. There's a lot of cross-collination between different groups when it comes to AI, because similar techniques can be

136
00:23:02.880 --> 00:23:07.180
STFC-RAL-CR03  R61: We use from one to another. And so in this case, we were just asking, can we…

137
00:23:07.930 --> 00:23:19.589
STFC-RAL-CR03  R61: That's a galaxy classification as a big project. We have lost a lot of data. We want to see if we do this, classify between… separate between spiral and elliptical galaxies using very similar

138
00:23:21.420 --> 00:23:24.489
STFC-RAL-CR03  R61: Using a very similar, similar outlet, outlet.

139
00:23:27.650 --> 00:23:35.879
STFC-RAL-CR03  R61: Okay, so one of the other things that people… so this is not from my group, but from one another. I think this is, CERN, a lot of certain people involved in this one.

140
00:23:36.100 --> 00:23:44.300
STFC-RAL-CR03  R61: One of the other things people have looked at is, you know, when it comes to drawing these, let's say, boundaries between groups, well, we could do this for

141
00:23:44.610 --> 00:23:51.179
STFC-RAL-CR03  R61: standard model versus non-stand models? Can we… can the quantum algorithm learn something from the data?

142
00:23:51.200 --> 00:24:06.100
STFC-RAL-CR03  R61: Can it learns some correlation or something or other that we weren't… that the classical one wasn't… isn't privy to. So the interesting thing with pharmacists is that, you know, we're dealing with inherently quantum processes, but the data that we have is classical.

143
00:24:06.480 --> 00:24:16.009
STFC-RAL-CR03  R61: So, it's quantitative processes, classical data, and then you're using either classical or quantitative processing to process that data. So, one of the things that people are…

144
00:24:16.510 --> 00:24:25.650
STFC-RAL-CR03  R61: What I need to ask right now is how many of the… how much of the correlation, the quantum correlation, are still present in that classical data that we could exploit with a… with a quantum machine?

145
00:24:26.170 --> 00:24:26.930
STFC-RAL-CR03  R61: And…

146
00:24:26.970 --> 00:24:43.449
STFC-RAL-CR03  R61: there has been a freezing interest in looking for new physics in this way, and, you know, as you know, after the Higgs boson discovery, this being a bit of a desert, we haven't yet seen anything. Some of our most compelling, well-motivated theories for what could come beyond the sound model haven't yet come to pass.

147
00:24:43.450 --> 00:24:52.769
STFC-RAL-CR03  R61: It could be that they're just around the corner, you know, the famous phrase of the sous vide and everything else just being around the corner, but we are now looking at

148
00:24:52.780 --> 00:24:55.889
STFC-RAL-CR03  R61: model agnostic signature. So we're trying to ask, you know.

149
00:24:56.080 --> 00:25:14.210
STFC-RAL-CR03  R61: Can we build an algorithm that just tells us, look, some… just flags an event and says, this does not look like a strong model. We don't know what it is. We're trying not to have this sort of top-down approach, but more of a bottom-up one, where we remain pretty much agnostic to the new physics thing that could be creeping up in our data.

150
00:25:14.290 --> 00:25:24.650
STFC-RAL-CR03  R61: So there has been increased interest in what they call unsupervised machine learning, where, you know, machine… you're sort of training it on data, data rather than your, your sort of simulation.

151
00:25:24.940 --> 00:25:28.989
STFC-RAL-CR03  R61: We're not telling them what to look for, but just to flag anything that looks anomalous.

152
00:25:29.960 --> 00:25:48.060
STFC-RAL-CR03  R61: So, in this particular case, they trained their algorithm. This was a quantum-enhanced, if you like, algorithm. This was trained on standard model processes, and the idea was to detect any outliers that may have physics that could potentially come from nuclear use. So, in this case, they looked at

153
00:25:48.540 --> 00:25:54.100
STFC-RAL-CR03  R61: just… it was just jet data, so it was just sort of QCD.

154
00:25:55.700 --> 00:26:06.040
STFC-RAL-CR03  R61: So here's a sort of full format of what they were doing. They have their features over here, and they were… in particular, they were interested in looking at

155
00:26:06.520 --> 00:26:15.519
STFC-RAL-CR03  R61: unprocessed… the raw… the more raw your data is, the better it was, in the sense that they want to preserve as many correlations as they can. So they were looking at

156
00:26:15.560 --> 00:26:29.009
STFC-RAL-CR03  R61: features of the data over here. This was, I think, two jets, and they looked at the hundreds, the top 100 PT particles constituting those two jets, and then there… the delta eta, delta phi, and PT of those… of those jets.

157
00:26:30.610 --> 00:26:37.960
STFC-RAL-CR03  R61: of those particles constituting their data. Then that data was compressed using a… a…

158
00:26:39.430 --> 00:26:55.660
STFC-RAL-CR03  R61: autoencoder format. So here, you're basically trying to reduce the dimensionality of your data, to some latent space. So this tries to learn all of the savior features of your dataset, and, you know, compress it down into some 4, 6, 8, or so parameters.

159
00:26:56.720 --> 00:27:12.779
STFC-RAL-CR03  R61: And then they were using that latent space to do non-next collection. So, there were several different types of algorithms that they looked at, clustering algorithms, and then the kernel machine, which is what I described before, where you're trying to draw some, complex decision boundary between your data.

160
00:27:14.630 --> 00:27:30.870
STFC-RAL-CR03  R61: So these were the results for… they looked at three signals, so this was one of them, was this narrow radical gap route tonic that came to two Wosomes. Another was a scalar boson that came to Higgs and Z, so you had three Z to the final state.

161
00:27:30.870 --> 00:27:48.240
STFC-RAL-CR03  R61: And a third one was a broad gravitational. So these have different… differing levels of complexity in terms of how similar they look to your standard model. So, of course, the more similar they are to what you expect in some models supposed to be, the more difficult it is to separate, between them.

162
00:27:48.320 --> 00:27:55.699
STFC-RAL-CR03  R61: And they… so these are… so… These different colors here for…

163
00:27:56.090 --> 00:28:10.290
STFC-RAL-CR03  R61: the very complicated blocks are here, because you have all three of these processes being shown. So if we just look at this one, so, you know, they had a look at, how many training samples do you need, and as you might expect, the larger the training side, the better the performance.

164
00:28:10.290 --> 00:28:16.670
STFC-RAL-CR03  R61: They looked at the space, the laser space dimension, so how much can you compress your data down before you can

165
00:28:16.670 --> 00:28:26.439
STFC-RAL-CR03  R61: You know, before you can still actually get a good performance out of it. And for the quantum performance, the best one was that you needed the latent dimension greater than 4.

166
00:28:26.780 --> 00:28:45.229
STFC-RAL-CR03  R61: And then over here, you can fix the other two, and then just look at, what these three signatures were showing. For both the quantum case, which is in the solid one, and the classical case. And then here, you can see that the quantum seems to be outperforming the classical for pretty much all three of the signature.

167
00:28:46.330 --> 00:28:49.660
STFC-RAL-CR03  R61: So this was quite interesting,

168
00:28:50.700 --> 00:29:05.860
STFC-RAL-CR03  R61: Yeah, so this is quite interesting, because it does show us that they're particularly interested in seeing whether the quantum is picking up, the quantum algorithm is picking up on something that has spoken. So, we're interested in seeing just how much overlap there is between the

169
00:29:05.860 --> 00:29:10.759
STFC-RAL-CR03  R61: The things that have been identified by the classical algorithm versus the ones by the quantum algorithm.

170
00:29:13.460 --> 00:29:23.249
STFC-RAL-CR03  R61: So another thing we looked at was the hot-on shower. So this is the step that comes in after the hard interaction. So hard interaction, of course, from how large

171
00:29:23.380 --> 00:29:40.280
STFC-RAL-CR03  R61: It's characterized by these large, momentum transfer, high energy, short distances, and after that, you have the parking shower, which takes… it evolves that scale of the heart interaction down to the hybridization scale, and so it's dominated by the loss of radiation of quarks and gluons.

172
00:29:40.280 --> 00:29:44.059
STFC-RAL-CR03  R61: And it's altogether quite a messy affair.

173
00:29:45.080 --> 00:29:48.029
STFC-RAL-CR03  R61: So, we looked at a…

174
00:29:48.340 --> 00:29:53.599
STFC-RAL-CR03  R61: quantum algorithm for this in one of our previous papers. So this was a quantum shower algorithm.

175
00:29:53.730 --> 00:29:55.860
STFC-RAL-CR03  R61: And we found that, you know.

176
00:29:56.000 --> 00:30:16.449
STFC-RAL-CR03  R61: For the classical case, you have to manually keep track of each individual shower history. So, at every point in this process, you have a proton, you know, a gluon giving you forks, and then a radiating again, etc. And so you have to physically… you have to keep track of each individual shower history and store it in a physical memory device for the classical case.

177
00:30:16.550 --> 00:30:29.969
STFC-RAL-CR03  R61: For the quantum algorithm, we saw that you could actually do this, calculate all possible shower histories pretty much simultaneously, so we were able to simulate a two-step, just two steps of that bottom shower process with about 31 units.

178
00:30:30.500 --> 00:30:50.450
STFC-RAL-CR03  R61: And we have to… for all of these, and this is true, you know, across the board, people have to simplify the problem down to some very toy, cookie model, so this is not at all close to what we're doing, with that, you know, classically. So these are very, very, very simplified models, because that's all we can work with right now with the sales machine that we have access.

179
00:30:52.880 --> 00:30:55.230
STFC-RAL-CR03  R61: But then we improved this algorithm.

180
00:30:55.340 --> 00:31:05.669
STFC-RAL-CR03  R61: And I'm going to… we use a quantum walk to describe the algorithm. So, before I tell you what a quantum walk is, so for a classical walk, you will start off at some

181
00:31:05.810 --> 00:31:16.560
STFC-RAL-CR03  R61: position in space over here. You would use a coin flip to determine whether… which direction you should move. So if you get a head, you move left, if you get a tail, you move right.

182
00:31:17.010 --> 00:31:31.039
STFC-RAL-CR03  R61: And then you do this all the way, and this is often called a drunkard swap, because you basically end up… the position that you end up in is very close to your initial position. So the more likely scenarios are, you're gonna end up somewhere over here, and the very unlikely scenarios are.

183
00:31:31.310 --> 00:31:41.600
STFC-RAL-CR03  R61: Further away from the center over here. So the average distance of the walker from the initial position scales as a square root, and our time, and it's the number of steps you've taken.

184
00:31:43.360 --> 00:31:58.030
STFC-RAL-CR03  R61: For a quantum walk, now, you have a quantum walker, which means that the quantum walker can be in a state. You also now have a quantum coin, so the quantum coin doesn't say, you know, heads or tail, it's in some superposition of heads or tail.

185
00:31:58.150 --> 00:32:07.619
STFC-RAL-CR03  R61: And so the quantum walker moves out in this kind of… in this kind of way. And the place where the quantum walker ends up is…

186
00:32:09.160 --> 00:32:19.889
STFC-RAL-CR03  R61: is this orange line right here. You can see that whereas a quantum… whereas a classical walker was expected to end up very close… the brokerage walker, you're expected to end up very close to your initial starting position.

187
00:32:19.900 --> 00:32:30.779
STFC-RAL-CR03  R61: The quantum walker is actually… Walker is quadratically faster than the astral one, so you can prove these, you know, extremities that you weren't able to… to get before.

188
00:32:31.780 --> 00:32:35.639
STFC-RAL-CR03  R61: So we looked at whether you could do something like this.

189
00:32:36.280 --> 00:32:47.320
STFC-RAL-CR03  R61: Or a parking shower, because in some ways, it's a similar process. You have… well, not really, but, you know, you can… you're sort of emitting something in each step of the process, and so we looked at whether you could do this.

190
00:32:47.590 --> 00:33:03.269
STFC-RAL-CR03  R61: So we designed a part-time shower, by reformulating it at a quantum walk in two dimensions. So, at each point flip, it basically tells you, are you going to make an… is there an emission or not? What's the probability of emitting a few months of walks?

191
00:33:03.390 --> 00:33:05.310
STFC-RAL-CR03  R61: And depending on the…

192
00:33:05.420 --> 00:33:16.429
STFC-RAL-CR03  R61: the… what that point clip tells you, then you either move in the X direction or the Y direction. So it's a two-dimension, because we're gonna connect either bruons or… or clock pairs.

193
00:33:16.550 --> 00:33:23.740
STFC-RAL-CR03  R61: So with this approach, we were able to simulate 31 steps of the parking shower using only 16 qubits.

194
00:33:24.440 --> 00:33:35.409
STFC-RAL-CR03  R61: So, in the previous algorithm, we could only do… in the previous quantum algorithm, we could only do two steps with 31, and now we're able to do a dramatically increase the number of steps with almost half the number of qubits.

195
00:33:35.640 --> 00:33:39.799
STFC-RAL-CR03  R61: So… The main point of showing this is that

196
00:33:40.020 --> 00:33:57.520
STFC-RAL-CR03  R61: Previously, we were doing what is, you know, what you could call a translational approach. You were basically taking a classical algorithm and just trying to port it over to a quantum device, which is probably not the best way to go about doing things, because these devices are so different, so, so, sort of, you know, dissimilar to each other.

197
00:33:57.700 --> 00:34:08.140
STFC-RAL-CR03  R61: That when we took a more, if you like, quantum-centric approach, where we thought about, you know, just, just erase everything we've done classically, and just think about how we might

198
00:34:08.139 --> 00:34:25.399
STFC-RAL-CR03  R61: we formulated this for a quantum device, we were able to do much better. So, I think there is going to be… have to be some kind of conceptual shift in how we think about putting our problems onto a quantum machine. Are you simply porting over our class algorithm is not going to be the most efficient way of putting that, right?

199
00:34:30.820 --> 00:34:47.269
STFC-RAL-CR03  R61: So one of the other things that we looked at, which might be of, interest also to people over here, was we looked at a neutrino dataset. This was a liquid argon TPC, so we wanted to see if we could classify track-like versus shower-like qualities in liquid argon.

200
00:34:47.440 --> 00:34:54.470
STFC-RAL-CR03  R61: This, of course, would help with improving the freedom of effort reconstruction in experiments like microboom and Dune.

201
00:34:54.810 --> 00:35:02.260
STFC-RAL-CR03  R61: We used quantum machine learning techniques, again, to classify them using detector hit data, so you see a typical neutrino,

202
00:35:02.750 --> 00:35:05.500
STFC-RAL-CR03  R61: Well, the screen and traction event over here, you have some

203
00:35:05.640 --> 00:35:13.580
STFC-RAL-CR03  R61: shower deposits from the photons, and then you have two tracks, tracked-like objects here, the muon and proton.

204
00:35:14.450 --> 00:35:31.830
STFC-RAL-CR03  R61: For this study, we used simulated microboom data and a synthetic data set that was developed so that we could vary the opening angles. So we just had one for primary vertex, and we had a muon and an electron coming out of that primary vertex, and we varied the opening angle just so we could

205
00:35:32.030 --> 00:35:44.400
STFC-RAL-CR03  R61: sort of dial up or down the complexity of the problem. So the closer they were together, the more difficult they were going to be to differentiate, and the further apart they were, the easier it was going to be. So this was a synthetic data set that we used.

206
00:35:45.610 --> 00:35:54.269
STFC-RAL-CR03  R61: So, we used a… so for classical CNN, so these are classical convolutional neural nets, so these use convolutional filters, which essentially

207
00:35:54.650 --> 00:35:55.530
STFC-RAL-CR03  R61: I guess.

208
00:35:57.870 --> 00:36:04.050
STFC-RAL-CR03  R61: Yes, it would essentially slide across the image and extract local spatial features from that image pad.

209
00:36:04.330 --> 00:36:12.819
STFC-RAL-CR03  R61: So in this case, we replaced these filters with a small quantum leak circuit. So each image patch over here is then encoded into qubits.

210
00:36:12.820 --> 00:36:28.100
STFC-RAL-CR03  R61: We apply a series of imagery transformations, and then the outputs measure the features from that little image patch, and then those… those features are then fed into a classical classifier, which tells you whether that looks like a shower or a really…

211
00:36:28.500 --> 00:36:30.420
STFC-RAL-CR03  R61: A shower or a track.

212
00:36:31.370 --> 00:36:47.680
STFC-RAL-CR03  R61: We also tried to add some symmetry to the problem, symmetries, the rotational symmetry over here, so rotating the input by 90 degrees, should give you equivalent output. So you can see, or here, you know, just, in terms of what it looks like, the pixel

213
00:36:48.850 --> 00:36:58.230
STFC-RAL-CR03  R61: of the tracks or showers for a 5x5 image size, 9x9, 21 by 21, etc. So, the smaller…

214
00:36:58.590 --> 00:37:07.689
STFC-RAL-CR03  R61: the smaller your pack size, the less contextual information you're getting to your quantifier. It doesn't have enough information to be able to classify.

215
00:37:07.690 --> 00:37:22.710
STFC-RAL-CR03  R61: But if you give it too much information, if you give… if the pass sign is too large, then there's a lot of irrelevant features, or features that are, you know, not relevant at all for the problem that's being classified, so it also doesn't do so well, and it will degrade that performance, because you're basically adding noise.

216
00:37:23.440 --> 00:37:26.509
STFC-RAL-CR03  R61: So there is a, sort of, optimal, image size.

217
00:37:27.070 --> 00:37:39.519
STFC-RAL-CR03  R61: So these were the results, these were the learning capability of various models that we looked at. These… I'm not… I won't go into details about what some of these are, but the CNN is…

218
00:37:39.760 --> 00:37:45.779
STFC-RAL-CR03  R61: this one here, this is your classical neural net. There is a… so…

219
00:37:46.090 --> 00:38:00.850
STFC-RAL-CR03  R61: benchmarking… how do you benchmark against possible one? It's an open question right now, in the sense that you could take your best possible machine and look pastel algorithm, and there's no way that consumer is going to be able to beat that. And so, what we've done is to look at

220
00:38:01.190 --> 00:38:19.390
STFC-RAL-CR03  R61: benchmarking our quantum algorithms against classical algorithms that have a similar number of parameters. So that big… there's… it's more of an actual comparison. So here's… that's what we're doing when we're comparing these two over here, and then this one that's outperforming all, you know, both of them, is a deep

221
00:38:19.880 --> 00:38:29.409
STFC-RAL-CR03  R61: classical neonat. So this has 100 times the number of representation, which is something that we can't really do. We can't construct a quantum algorithm that can do that, because we don't have the… we don't have the ability.

222
00:38:30.120 --> 00:38:36.259
STFC-RAL-CR03  R61: So, you can see that the quantum algorithm slightly outperforms is its apples-to-apples comparison classical one.

223
00:38:37.090 --> 00:38:49.319
STFC-RAL-CR03  R61: We found that the quantum models tend to generalize better with limited data, so in other words, we can… and this has been found in personal studies across the board as well, that they seem to be able to train

224
00:38:49.830 --> 00:38:56.159
STFC-RAL-CR03  R61: To generalize with limited data, meaning that they can… they train on smaller sample sizes, so that after

225
00:38:56.310 --> 00:39:03.690
STFC-RAL-CR03  R61: they basically learn faster. So, you know, you give them a small sample size, and they're able to actually learn from that better than a classical one would.

226
00:39:03.860 --> 00:39:14.470
STFC-RAL-CR03  R61: We found that added symmetry to the problem gives you very minor gains, so there wasn't really any clear advantage, where this could just be that we haven't yet explored some of the higher order symmetry groups.

227
00:39:14.570 --> 00:39:16.640
STFC-RAL-CR03  R61: And,

228
00:39:17.360 --> 00:39:30.719
STFC-RAL-CR03  R61: And so for future work, one of the things we wanted to look at was to see if we could do something like hybrid quantum classical models and integrate them into Dune's reconstruction pipeline, especially Pandora.

229
00:39:30.720 --> 00:39:45.050
STFC-RAL-CR03  R61: So, which has this modular setup, which is… lends itself quite nicely to doing something… doing something like this. And also for separating shower-like particles from different particles, so whether it can identify whether one particular shower is from a photon or electron.

230
00:39:46.760 --> 00:39:50.680
STFC-RAL-CR03  R61: Alright, so this is… So, I think the next…

231
00:39:50.910 --> 00:40:00.649
STFC-RAL-CR03  R61: two, results that I'm going to present are slightly in a slightly different capacity. So this is now turning things around and asking.

232
00:40:01.120 --> 00:40:12.730
STFC-RAL-CR03  R61: Before this, we were looking at what can quantum computers do for particle physics, and now we're asking, what can particle physics do for quantum computing, or quantum information, or quantum mechanics?

233
00:40:12.800 --> 00:40:31.989
STFC-RAL-CR03  R61: So one of the things that people have looked at is, can we use high-energy processes, especially particle physics at the plate laboratory, for testing quantum foundations in ways that we haven't tested before, and that's a new energy regime? So this is… people have looked at testing belt inequalities at, I should say, this was proposed, this was a theory paper.

234
00:40:32.240 --> 00:40:41.819
STFC-RAL-CR03  R61: So they propose using pairs of top quarks produced at the LSC to code quantum entanglements at high energy where they haven't been explored.

235
00:40:42.110 --> 00:40:55.149
STFC-RAL-CR03  R61: So the top anti-top system lends itself quite nicely to something like this. It's a natural two-qubit state, so there's… the top fork has a spin-half, it behaves like a qubit, and a pair of them behaves like a… behaves like two qubits.

236
00:40:55.390 --> 00:41:03.180
STFC-RAL-CR03  R61: It's also nice that it decays before it hagonizes. All of the spin information is then printed on its decayed products. If you measure the… the…

237
00:41:03.540 --> 00:41:08.500
STFC-RAL-CR03  R61: the decay products, you have some information about the top screen.

238
00:41:08.680 --> 00:41:21.170
STFC-RAL-CR03  R61: And so it was using spin correlations off the top and anti-top, that they were able to reconstruct this, a correlation matrix that enabled them to probe and tackle it.

239
00:41:21.480 --> 00:41:27.650
STFC-RAL-CR03  R61: So they constructed an observable from the correlation matrix, and, to test

240
00:41:28.420 --> 00:41:33.429
STFC-RAL-CR03  R61: I keep saying they did this, but actually this is the theory paper suggesting that you could

241
00:41:35.370 --> 00:41:37.480
STFC-RAL-CR03  R61: So over here, they've…

242
00:41:37.640 --> 00:41:53.319
STFC-RAL-CR03  R61: They looked at what region of phase space could you see entitlement for easily already. And they found two particular regions. So this is the scattering angle between the top and anti-top. This is the invariant mass of the top there. So anything above

243
00:41:53.630 --> 00:41:59.020
STFC-RAL-CR03  R61: M1 plus M2 has to be greater than M1 in order for you to…

244
00:41:59.320 --> 00:42:06.899
STFC-RAL-CR03  R61: In order for you to see entanglement. So, you can see that you see this near the top mass.

245
00:42:07.740 --> 00:42:18.159
STFC-RAL-CR03  R61: Top, anti-top mass threshold over here, and also in the top right corner over here, where the scattering angle is large, and your top pair is very high energy.

246
00:42:18.260 --> 00:42:27.739
STFC-RAL-CR03  R61: So these are the only two regions, that show anything above 1, but of these, it's only the high energy region that shows some clean, strong valve violation.

247
00:42:30.220 --> 00:42:38.729
STFC-RAL-CR03  R61: And this is where both of the production mechanisms of the top, so the blue one, blue one, and the align in such a way to the maximally entangled square space.

248
00:42:38.730 --> 00:42:51.460
STFC-RAL-CR03  R61: So they proposed that with the existing LSC round 2 data, you could get something like a 4-sigma significance, sorry, a 98% confidence level, and with LSC round 3, you could get a 4-sigma significance.

249
00:42:52.650 --> 00:43:10.889
STFC-RAL-CR03  R61: So, Atlas and CMS actually did something to probe that entanglement using top pair. So, Atlas, again, they used… for the measure of entanglement, they used this variable D, which is essentially a proxy that you determine from the spin correlation of the… of the top four decay products.

250
00:43:11.220 --> 00:43:20.450
STFC-RAL-CR03  R61: So Atlas looked at this in a single region, which is close to threshold, so this is the region where you expect the top pairs to be maximally entangled.

251
00:43:20.650 --> 00:43:27.859
STFC-RAL-CR03  R61: And then had validation regions in the higher top mass, top and top invariant mass.

252
00:43:29.140 --> 00:43:34.690
STFC-RAL-CR03  R61: They used the full, I think this is the 140 inverse quantum miles of data.

253
00:43:35.180 --> 00:43:52.969
STFC-RAL-CR03  R61: they found… so this is the entanglement limit, and they were well below that, which means that they… they're basically more than 5 sigma standard deviation from a scenario without entanglement. So they… this was… they claim that this was the first observation of entanglement in a Hawk-ante-hawk pair.

254
00:43:53.150 --> 00:44:06.699
STFC-RAL-CR03  R61: Cms also did the same measurement, with a bid likelihood fit, and they got, a 4.7 sigma, solution of, top block entanglement.

255
00:44:08.400 --> 00:44:18.809
STFC-RAL-CR03  R61: And then just, I think a few days ago, so I was like, well, this is really, really exciting that we can actually use high-energy particle colliders as a lab to test high-energy foundation.

256
00:44:18.810 --> 00:44:30.759
STFC-RAL-CR03  R61: But clearly, there is a lot of discussions being had in the theory community over here. This is… these are papers that I just found a couple of days ago from people quite recently, I think it's in July and October, so people have

257
00:44:31.280 --> 00:44:51.059
STFC-RAL-CR03  R61: question whether we are actually testing anything here, whether we can actually test Belden equality or interim at the LSE. They've said that you can construct some local hidden variable theory that can explain your beta pretty well without having to invoke, or, you know, without it being a test of Belden equality.

258
00:44:51.130 --> 00:45:07.739
STFC-RAL-CR03  R61: And then we had another… this was another paper that said, that wasn't probably quite as negative. It said that, yes, you can't rule out local realism with these… with these experiments, but that you can probably measure. There are enough quantities for everybody to be able to say something about impact

259
00:45:08.110 --> 00:45:11.979
STFC-RAL-CR03  R61: So, I don't know yet what the, I mean, what the consensus is,

260
00:45:12.270 --> 00:45:15.789
STFC-RAL-CR03  R61: And… but there definitely seems to be some debate about whether we can

261
00:45:15.920 --> 00:45:21.479
STFC-RAL-CR03  R61: Whether the things we are seeing are actually what we say… what we think they are.

262
00:45:22.550 --> 00:45:34.079
STFC-RAL-CR03  R61: So another one is, so this is where quantum information consents have now been imported over into quantum physics, and, it's interesting, again, because we're looking at

263
00:45:34.220 --> 00:45:52.869
STFC-RAL-CR03  R61: The two, you know, we think that the two key phenomenon, which is superposition and entanglement, might be… are what distinguish quantum from classical, but actually, we found that just having entanglement in your system, just having entangled states in your quantum algorithms is not enough for… to give you a speed that runs to your classical computing.

264
00:45:52.970 --> 00:45:58.450
STFC-RAL-CR03  R61: It's not in a sufficient condition by itself to be able to play in Sweden.

265
00:45:58.560 --> 00:46:02.709
STFC-RAL-CR03  R61: And for any number of qubits, you actually need to have…

266
00:46:02.930 --> 00:46:08.010
STFC-RAL-CR03  R61: You actually need to have states that are what we call non-stabilizer states.

267
00:46:08.280 --> 00:46:12.739
STFC-RAL-CR03  R61: And, you know, the sort of quantum information parallels with all that magic.

268
00:46:13.160 --> 00:46:22.720
STFC-RAL-CR03  R61: So you need to have some degree of magic in your class… in your quantum algorithm to ensure that it's not classically simulable. So the more magic you have in your

269
00:46:22.960 --> 00:46:31.490
STFC-RAL-CR03  R61: Algorithm, the less easily classically significant legal it is, and therefore, the more important it is for being able to just claim advantage over a classical algorithm.

270
00:46:32.010 --> 00:46:33.520
STFC-RAL-CR03  R61: So,

271
00:46:34.770 --> 00:46:53.660
STFC-RAL-CR03  R61: So these were, so the theory paper, and they've asked, does nature produce magic saves at high energies? And they, in particular, looked at whether you could do this with, again, with, with top bark. So how much of this… so this is, you know, you can think of magic as being this property that's very,

272
00:46:54.360 --> 00:47:04.759
STFC-RAL-CR03  R61: We really want things to have magic, because the more magic that they have, the less easy it is for us to… for us to assimilate them, and the more valuable they are for quantum computing.

273
00:47:05.510 --> 00:47:22.369
STFC-RAL-CR03  R61: So they looked at just how much magic does nature produce in terms of the top box, and they… this is… these plots here are for the scattering angle in two different top production channels, so it's QP bar and yuan Yuan.

274
00:47:22.480 --> 00:47:24.890
STFC-RAL-CR03  R61: And anything above…

275
00:47:24.990 --> 00:47:43.020
STFC-RAL-CR03  R61: So anything, I think, in the… anything above zero tells you that, you have some degree of, magic in your system. So you say… see that across the base space, there is quite a considerable degree of, you know, zero magic regions across that, base space. And the ones where you get

276
00:47:43.560 --> 00:47:59.820
STFC-RAL-CR03  R61: interestingly, the maximal entanglement, where you expect the two top blocks to be maximally entangled, which is near threshold, does not always align with having maximal magic. So this is… these are different… there's different computational measures that,

277
00:48:00.660 --> 00:48:06.380
STFC-RAL-CR03  R61: As it measures through computational potential, not necessarily the amount of quantum correlation between our system.

278
00:48:06.430 --> 00:48:24.129
STFC-RAL-CR03  R61: And one of the future directions that they've identified is whether we could use this as a new observable to probe physics beyond examples. Can we incorporate that into our, you know, new design signatures and see whether it's a valuable resource for looking at physics beyond, beyond what we've seen so far?

279
00:48:26.600 --> 00:48:27.709
STFC-RAL-CR03  R61: Yeah, web done.

280
00:48:28.310 --> 00:48:41.000
STFC-RAL-CR03  R61: So yeah, I'm just on my last slide. So yeah, it's a very… the meeting is a very rapidly developing field. I mean, every, every other day or so, there's something in the news about something over there.

281
00:48:41.000 --> 00:48:50.959
STFC-RAL-CR03  R61: And there seems to have potential applications in a variety of areas. I think some of the most… some of the killer apps of the technology you probably haven't identified yet, and part of the whole

282
00:48:50.960 --> 00:49:10.560
STFC-RAL-CR03  R61: enterprise of, you know, doing this right now, doing these exploratory studies across different experiments, i.e. not tying yourself down to, you know, CMS Atlas, or you're sort of going across, you know, physics, etc, is to see… is to match use cases to algorithms. Like, where… where are we likely to see the most gains, and with what kinds of algorithms?

283
00:49:10.690 --> 00:49:27.460
STFC-RAL-CR03  R61: And also being able to have this conceptual paradigm shift in how you think about your problem when you bring it on a planning device. So, like I said, I showed you with the cotton shower example, you know, simply porting over your classical algorithm might not… is likely not going to give you the most, the most benefit.

284
00:49:27.550 --> 00:49:36.150
STFC-RAL-CR03  R61: The current machines are very small, they're noisy, they're not very good for, you know, anything, doing anything real, real, real scale.

285
00:49:36.520 --> 00:49:41.000
STFC-RAL-CR03  R61: But they are a very good testbed for demonstrating these kinds of proof-of-principle studies.

286
00:49:41.190 --> 00:49:55.019
STFC-RAL-CR03  R61: And there's also… I showed some of the, sort of, bi-direction exchange between these two groups of quantum community, or quantum information and partial physics, and that discussion you saw is still ongoing, so we'll see… we'll see how that evolves.

287
00:50:02.140 --> 00:50:04.589
STFC-RAL-CR03  R61: Questions? Yeah, Joe?

288
00:50:04.740 --> 00:50:21.130
STFC-RAL-CR03  R61: So I'm really interested in the, unsupervised learning capabilities, of quantum computing, but, I mean, given that it, it can learn really quickly, like, a lot quicker than, than, you know, like a classical convolutional neural network of similar size.

289
00:50:21.200 --> 00:50:26.729
STFC-RAL-CR03  R61: I mean, how… how are we sure in the tests that you were showing that it's not, like.

290
00:50:26.860 --> 00:50:30.020
STFC-RAL-CR03  R61: Overfitting something and detecting something that's

291
00:50:30.230 --> 00:50:45.540
STFC-RAL-CR03  R61: like, just kind of a random fluctuation that it thinks is real? Like, how… is there any risk of, like, when we start using quantum unsupervised living algorithms, doing this bottom-up approach of trying to find anomalies, is it going to find something spurious?

292
00:50:47.320 --> 00:51:05.559
STFC-RAL-CR03  R61: Yeah, that's, like, what's the… Yeah, I mean, so a lot of the things… so with overfitting, there are… we do do all of the, sort of, kind of parameter optimization and everything. We do do all the checks for overfitting that you would do with economical, you know, classical… classical machine learning approach. So those tests are done, but I guess it needs…

293
00:51:05.670 --> 00:51:10.410
STFC-RAL-CR03  R61: And hence the, you know, they looked at 3 different signals as well.

294
00:51:10.630 --> 00:51:16.380
STFC-RAL-CR03  R61: And they quantum reproduced what they would expect of the three different signals in terms of

295
00:51:16.480 --> 00:51:34.779
STFC-RAL-CR03  R61: Here. In terms of… some of these signals are easier to… they're closer to your sound model background, so they're much more complex to separate, whereas the ones that are definitely very different. I think it's a broad route, that's quite different from the… from the sound model. So, quantum seems to reproduce the classical one pretty well, over here.

296
00:51:34.780 --> 00:51:46.469
STFC-RAL-CR03  R61: And, I mean, yeah, there will have to be, sort of, checks and balances to make sure that we don't… we can't just throw machinery like this at the data and expect it to… and expect it's just going to all work out. But…

297
00:51:46.520 --> 00:51:52.460
STFC-RAL-CR03  R61: So doing simulation studies like this are going to help us to be… have more confidence in whatever algorithm we come up with.

298
00:51:52.890 --> 00:52:02.319
STFC-RAL-CR03  R61: So, like, is this gonna be some… like, so, I mean, you've got… sorry if you don't want the quick follow-up. So the… so you have the classical…

299
00:52:02.440 --> 00:52:17.799
STFC-RAL-CR03  R61: It was here. And actually, while I'm at it, what's TPR and FPR? I… I mean, for our, it's all positive rating. Okay, okay, okay. But, yeah, single efficiency and background efficiency. Yeah, okay.

300
00:52:18.350 --> 00:52:32.620
STFC-RAL-CR03  R61: And so, one of the things that we might do, at least maybe at the beginning, is kind of having them run side by side, and say, okay, like, here's a, you know, a classical computer where we kind of know, like, we know the terrain a little bit better.

301
00:52:32.620 --> 00:52:41.560
STFC-RAL-CR03  R61: And then here's the quantum conflict, so the quantum computer can kind of reach a little bit further, so it's backed by something that is, you know, people are going to be less scared of.

302
00:52:41.830 --> 00:52:53.789
STFC-RAL-CR03  R61: And then you're, and then you're using quantum machinery to sort of, like, extend from there. So it's kind of like a stuff without context, I think. Yeah, I think it might take a similar…

303
00:52:54.580 --> 00:53:13.360
STFC-RAL-CR03  R61: It might be told in a similar way to how we did with machine learning as well. I mean, people used to have this, we used to have two parallel analyses in the CMS and Atlas experiments. It was cut and count, and then you can use the gradient decision tree, or, you know, one of the decision boost tree, or something like that. You're always going to have a cut and count approach to

304
00:53:13.360 --> 00:53:26.739
STFC-RAL-CR03  R61: to complement it, because we weren't quite sure, because they were a bit more black boxy, and we weren't quite sure what they were doing, and so… and then we've had that, and many of the important analyses in Apple and CMS had both of them running alongside, and you had to make sure that your results matched up.

305
00:53:27.070 --> 00:53:31.570
STFC-RAL-CR03  R61: So it's likely that it's going to take a similar… have a similar approach there.

306
00:53:32.220 --> 00:53:33.530
STFC-RAL-CR03  R61: Well, we're spending…

307
00:53:33.810 --> 00:53:42.749
STFC-RAL-CR03  R61: Your support vector machine knew for track finding. I wasn't clear whether you were running a quantum algorithm on classical hardware, or you were actually running on a quantum machine.

308
00:53:43.430 --> 00:53:47.800
STFC-RAL-CR03  R61: Interesting question, because for a lot of these, I think we have

309
00:53:48.740 --> 00:53:51.849
STFC-RAL-CR03  R61: The quantum machine that we have access to on, like, 5GBTs.

310
00:53:51.850 --> 00:54:13.730
STFC-RAL-CR03  R61: So we do run on the plant machines, but we have to pare down the problem to such a simplified toy model that, you know, we essentially just do it just to check that things are looking as we would expect. But for most of these, we are still using a classical simulator. So this is a classical, 32 cubic simulator is the highest, machine that we have access to.

311
00:54:14.830 --> 00:54:16.870
STFC-RAL-CR03  R61: Yeah, and then we… then we…

312
00:54:17.020 --> 00:54:23.050
STFC-RAL-CR03  R61: We also do it for a 5-foot device, but mostly just to check, I think, that we can do it.

313
00:54:23.250 --> 00:54:31.999
STFC-RAL-CR03  R61: I'm afraid there's a follow-up. I don't suppose you noticed there was a paper last Thursday doing the same thing for dietic studies, and I couldn't work out whether they were using CLIC.

314
00:54:32.060 --> 00:54:48.460
STFC-RAL-CR03  R61: Right, right. Okay. Yeah, and most of the people right now, they're working on the simulator, because they don't have access to devices that are… I think the publicly accessible ones, that's all we have access to right now, are by IBM, and they're, like, 5-7 cubits or so.

315
00:54:49.950 --> 00:55:08.689
STFC-RAL-CR03  R61: But yeah, one of the things I've been trying to tell my students, even if they are using a simulator for the entirety of their… of the paper, is at the end to actually draw and work on an actual machine, because it's the difference between night and day in some ways, because it's different between running with simulated data and actual data, that it can be, you know.

316
00:55:08.690 --> 00:55:19.510
STFC-RAL-CR03  R61: It can change very… it can change a lot, because you have to deal with noise profile when, you know, the machine is calibrated every single day, and you just have to take the noise profile for that particular day, and…

317
00:55:19.510 --> 00:55:25.549
STFC-RAL-CR03  R61: implemented into your algorithm, so it's a different ballgame, and yeah, we have to know how to do it.

318
00:55:26.930 --> 00:55:37.640
STFC-RAL-CR03  R61: Yes. More of a comment, so you mentioned the theory paper on Magic, the LFC. Quite recently, CMS has actually followed that paper and put it out a path to doing… Right, okay.

319
00:55:37.720 --> 00:55:44.449
STFC-RAL-CR03  R61: It's a little bit… the interpretation's a little bit more difficult than things like entanglement, because

320
00:55:44.460 --> 00:56:01.020
STFC-RAL-CR03  R61: It's an inherently basis-dependent quantity, so if you redefine the spin axes, which is used to make the measurement, you get a different magic. Right. So when you have this mixture of all sorts of quantum states in your TT bar, there's a… it makes it very difficult to say what is the relevance of the measurements.

321
00:56:01.140 --> 00:56:10.050
STFC-RAL-CR03  R61: So, we've actually got a follow-up paper coming soon. Magic actually isn't in that because of this difficulty interpretation, but that should be coming out in the next…

322
00:56:10.100 --> 00:56:27.709
STFC-RAL-CR03  R61: next few days. Right. Also on the high… high PT, phase space of the TT car system, CMS put out, a measurement of the, entanglement there, and has an observation also in the hunt. All right, okay. Okay, it's not just close to threshold, but now also high.

323
00:56:27.980 --> 00:56:42.029
STFC-RAL-CR03  R61: Yeah, sorry, I had the same sort of question as Bill, actually. But… so on your… in your simulated systems, do you have any simulation of the errors that you might get?

324
00:56:42.300 --> 00:56:45.100
STFC-RAL-CR03  R61: In qubits, and how many error correction.

325
00:56:45.230 --> 00:56:47.300
STFC-RAL-CR03  R61: As part of the exercise.

326
00:56:47.400 --> 00:57:04.059
STFC-RAL-CR03  R61: there's no error correction, but there are… I mean, we do have some basics, like some readout error and things like that. But I wouldn't say that that machinery is particularly sophisticated right now, in the sense that, you know, yeah, so we incorporate readout errors, and…

327
00:57:04.060 --> 00:57:20.640
STFC-RAL-CR03  R61: numbers one or the other, but it's not, yeah, we don't have a full… it's not like, you know, I sometimes show these thoughts that come to CMS, you know, the first question in my mind when my students show up this, and I'm like, what's describing? So I think that, and it's very… it's become… it's difficult right now to be able to

328
00:57:20.750 --> 00:57:22.090
STFC-RAL-CR03  R61: get a full…

329
00:57:22.880 --> 00:57:31.800
STFC-RAL-CR03  R61: And so, also, since you're doing it through simulation, what's the sort of limit on the number of keywords that you can simulate in any of these?

330
00:57:32.130 --> 00:57:33.270
STFC-RAL-CR03  R61: And…

331
00:57:34.380 --> 00:57:52.549
STFC-RAL-CR03  R61: I mean, in terms of what we have access to, I think it's 32 qubits simulated. There are larger ones, I think there might be 50-something. But I know that, I think it's beyond, like, 60 or so qubits. I think we are going to be close to the point where you can't really simulate them anymore. I mean, I think what's the number? It's, like, what, 300 qubits? You need, like, more than

332
00:57:52.900 --> 00:57:59.519
STFC-RAL-CR03  R61: So the observable universe is going to be able to simulate that. So we are… we are getting close to that point where we won't be able to.

333
00:58:01.490 --> 00:58:03.119
STFC-RAL-CR03  R61: Are there any questions on Zoom?

334
00:58:03.740 --> 00:58:04.930
STFC-RAL-CR03  R61: You can't see it.

335
00:58:05.570 --> 00:58:10.450
STFC-RAL-CR03  R61: Sorry, maybe the most first, please.

336
00:58:11.920 --> 00:58:13.300
STFC-RAL-CR03  R61: You look great.

337
00:58:13.870 --> 00:58:19.939
STFC-RAL-CR03  R61: So, it's… it's a bit of a more general question about quantum computers. So, I was…

338
00:58:20.190 --> 00:58:26.850
STFC-RAL-CR03  R61: So we read a lot about computers, but I was a bit confused about how to actually understand the

339
00:58:26.980 --> 00:58:29.110
STFC-RAL-CR03  R61: performance of quantum computers.

340
00:58:29.340 --> 00:58:44.949
STFC-RAL-CR03  R61: So, can I sort of understand, like, the number of qubits, sort of like, the register bit size of classical, CPUs? For example, 32-bit, 64-bit processors? Would that be something similar to, like, a

341
00:58:45.090 --> 00:58:47.880
STFC-RAL-CR03  R61: 60 more qubit on the computer.

342
00:58:48.950 --> 00:58:55.369
STFC-RAL-CR03  R61: So before… so there has been some sort of shift in the language, how people talk about

343
00:58:55.590 --> 00:59:11.199
STFC-RAL-CR03  R61: the power plants and machines, because before, there was a lot of emphasis on keywords, but now we realize that, actually, you could have hundreds of keywords right there. If they're low… they're higher, they're noisy keywords, you don't really get anything out of them. You might as well just, like, anything into them, and you get…

344
00:59:11.380 --> 00:59:17.670
STFC-RAL-CR03  R61: rubbish in, rubbish out type of thing. So now there's entire… I don't think I have it in these slides, but there's entire

345
00:59:17.810 --> 00:59:37.249
STFC-RAL-CR03  R61: something called quantum building, which actually incorporates many more metrics into that, not just the number of qubits. So you need the number of qubits, you need the error rates, the amount of crosstalk between qubits. You only want qubits talking to each other that you want to talk to each other, you don't want crosstalk between qubits that are not involved in your computation.

346
00:59:37.430 --> 00:59:48.439
STFC-RAL-CR03  R61: What are the… there's, there's… yeah, IBM have come across… people have come up with different metrics, but one of the ones that IBM is using is plant volume, so this incorporates a lot more than just the remote units.

347
00:59:48.680 --> 00:59:58.379
STFC-RAL-CR03  R61: And in terms of how to compare that to FastLoon, I'm not sure that there's a good… I don't know how to… I don't know whether there's a dictionary between one and the other.

348
00:59:58.620 --> 01:00:05.300
STFC-RAL-CR03  R61: Yet. But number of tumors is not DMV. In fact, it's now considered not a great metric.

349
01:00:06.240 --> 01:00:20.699
STFC-RAL-CR03  R61: But, like, considering, like, what classical and quantum… at least, like, if you say, like, you don't consider that everything too much. Yeah. They're just basically, like, matrix operations for, it's…

350
01:00:21.240 --> 01:00:23.880
STFC-RAL-CR03  R61: So… Is it not?

351
01:00:26.280 --> 01:00:30.419
STFC-RAL-CR03  R61: I mean, yes, in a sense… well, I wanna…

352
01:00:31.560 --> 01:00:39.270
STFC-RAL-CR03  R61: I want to make sure that we, you know, it's, I think I showed it with about hundreds of,

353
01:00:39.420 --> 01:00:51.929
STFC-RAL-CR03  R61: So you need something around these many, and I think these were actually logical qubits to be able to do something, something rather than something in the real world.

354
01:00:52.190 --> 01:01:00.579
STFC-RAL-CR03  R61: I don't know how to… yeah, I'm not sure I can answer your question in terms of how do you compare that to a classical machine. At what point would we be able to do that?

355
01:01:00.740 --> 01:01:04.729
STFC-RAL-CR03  R61: Because they're, I think, Well, if we get to Universal.

356
01:01:04.970 --> 01:01:11.359
STFC-RAL-CR03  R61: active computation on me, but until we're there, I mean, while we're still dealing with this, I don't think there's going to be a clear…

357
01:01:12.510 --> 01:01:14.629
STFC-RAL-CR03  R61: Clear some go-between distribution.

358
01:01:16.200 --> 01:01:17.090
STFC-RAL-CR03  R61: Yep.

359
01:01:17.460 --> 01:01:24.040
STFC-RAL-CR03  R61: So… Ask me for a friend, but didn't you go to the quantum block again?

360
01:01:32.110 --> 01:01:33.000
STFC-RAL-CR03  R61: Yeah.

361
01:01:33.420 --> 01:01:34.220
STFC-RAL-CR03  R61: Out?

362
01:01:34.700 --> 01:01:46.410
STFC-RAL-CR03  R61: So, you've got a probability distribution in the classical sense, which makes very intuitive sense. How is there higher probability for the quantum walk to have

363
01:01:47.170 --> 01:01:51.810
STFC-RAL-CR03  R61: like, basically higher probability at the fringes. Like, it kind of looks like they got these, like, 1 over…

364
01:01:52.250 --> 01:01:59.630
STFC-RAL-CR03  R61: that, like… So you get some destructive interference over here? Oh, so, okay, so it's an interference thing. Yeah. Okay, gotcha. And then you have, yeah.

365
01:02:00.750 --> 01:02:06.239
STFC-RAL-CR03  R61: Okay, yeah, that'd be slow. Because it's both the walker is quantum and the coin is quantum, so the…

366
01:02:06.690 --> 01:02:08.549
STFC-RAL-CR03  R61: Okay, okay, gotcha.

367
01:02:10.500 --> 01:02:27.449
STFC-RAL-CR03  R61: James said, Yeah, I was going to say, setting aside places like this, but if you're a particular approaches kind of what history of all… you want, kind of, some of all possible account history is. You're talking about machine learning as a sponsor of machine learning. Where is… where do you think the gains are? Is it…

368
01:02:27.820 --> 01:02:32.030
STFC-RAL-CR03  R61: Where the fated space has to be high-dimensional, and you kind of want to be able to

369
01:02:32.290 --> 01:02:35.090
STFC-RAL-CR03  R61: Where we don't all simultaneously at the same time.

370
01:02:35.210 --> 01:02:42.480
STFC-RAL-CR03  R61: Is that where… where the… Where the bottom machine learning equipment is the most part, do you think?

371
01:02:43.630 --> 01:02:47.210
STFC-RAL-CR03  R61: People have seen…

372
01:02:47.510 --> 01:03:05.609
STFC-RAL-CR03  R61: I think in terms of the machine learning, especially in approaching the new physics problems with machine learning, people have seen the most benefit in that autoencoder approach, where you do reduce your data set down to some latent space, and then code the latency space directly for, you know, for information.

373
01:03:05.780 --> 01:03:16.770
STFC-RAL-CR03  R61: And I think one of the reasons… so it could be that that might… that's one of the advantages. And you want to have the data in as raw a format as possible, so you do, you know, minimal

374
01:03:16.770 --> 01:03:26.720
STFC-RAL-CR03  R61: classical post-processing of the data. So you're not tweaking any… so you're not getting rid of any correlation that might… point of correlation that might already linear data that your fat model doesn't exploit.

375
01:03:26.880 --> 01:03:41.259
STFC-RAL-CR03  R61: And one of the advantages, I think, of doing the latency space is that if you give out… if you're… if you're… the… you know, the data that you're reducing down, if that's raw data, just, you know, formatta of your particles, etc.

376
01:03:41.430 --> 01:03:43.729
STFC-RAL-CR03  R61: And with minimal tweaking.

377
01:03:43.840 --> 01:03:55.580
STFC-RAL-CR03  R61: then you might still be preserve enough information, and then if they've had some… if that's quite small autoencoder, it might be able to learn representations of that data.

378
01:03:55.600 --> 01:04:10.500
STFC-RAL-CR03  R61: So it's a different space rather than just better at exploring high-dimensional links. I mean, it is, it is better at exploring that high-dimensional space, but, it's very difficult to say what it might be doing that, yeah, yeah. Yeah.

379
01:04:11.460 --> 01:04:26.390
STFC-RAL-CR03  R61: So, it's sort of a very silly question, but can you, with the quantum machine learning, can you… does it have to be quantum to run inference on it as well, or can you run it… does it produce weights that you can run classically to evaluate it, or does it have to be in your quantity to evaluate it?

380
01:04:26.970 --> 01:04:42.920
STFC-RAL-CR03  R61: Well, we're doing… most of the… those approaches are hybrid, so you're doing either classical pre-processing or, you know, classical post-processing on that data. So you're doing… it's all the… it's more modular in some ways than… than doing everything in a… in a quarantine.

381
01:04:43.460 --> 01:04:45.009
STFC-RAL-CR03  R61: I'm gonna find tomorrow.

382
01:04:45.270 --> 01:04:50.189
STFC-RAL-CR03  R61: So you are doing stuff classically on that, on the results of your personality.

383
01:04:51.120 --> 01:04:55.200
STFC-RAL-CR03  R61: Okay, but you still need the other computer, not just the screen.

384
01:04:56.110 --> 01:05:00.010
STFC-RAL-CR03  R61: Just curious, Sid, by how easy it is, actually, to use this.

385
01:05:01.650 --> 01:05:08.970
STFC-RAL-CR03  R61: Sometimes we're doing the training classically, yeah. Okay, okay. Yeah. So it depends. Okay, thanks, thank you.

386
01:05:10.910 --> 01:05:15.390
STFC-RAL-CR03  R61: Are there any hands up on Zoom, or any other questions from the room?

387
01:05:18.320 --> 01:05:21.040
STFC-RAL-CR03  R61: If not, then let's thank you again.

388
01:05:26.260 --> 01:05:31.369
STFC-RAL-CR03  R61: So, we'll be heading for lunch in probably a bit in 7-8 minutes, so…

