WEBVTT 1 00:00:02.020 --> 00:00:03.979 Sadia Afrin Purba: Yeah. Share your screen. 2 00:00:05.140 --> 00:00:07.299 Sadia Afrin Purba: Otherwise, I won't be able to see slides. 3 00:00:10.760 --> 00:00:15.190 Sadia Afrin Purba: Okay, so thank you all for joining us today. We are here to… 4 00:00:15.460 --> 00:00:26.510 Sadia Afrin Purba: Towards for Sadia for the next hour and a half, of all of us. Now, this is Sadia's PhD preliminary exam. Sadia's in her third semester here at Temple. 5 00:00:26.810 --> 00:00:29.329 Sadia Afrin Purba: Pursuing a PhD. 6 00:00:29.870 --> 00:00:33.050 Sadia Afrin Purba: Dissertation area is probably going to be… 7 00:00:33.190 --> 00:00:39.570 Sadia Afrin Purba: Quantum entanglement and sheet learning type things, so it's going to be actually fairly close to what she's talking about today. 8 00:00:39.750 --> 00:00:44.619 Sadia Afrin Purba: But to remind everybody, this is a preliminary exam, not a dissertation proposal. 9 00:00:45.100 --> 00:00:46.120 Sadia Afrin Purba: Oh, okay. 10 00:00:46.360 --> 00:00:51.930 Sadia Afrin Purba: Savi's had a pretty productive time in our group. She's got a number of menace publications coming out this year. 11 00:00:52.240 --> 00:01:00.019 Sadia Afrin Purba: And she's so… Most of them evolving, around our work in machine learning. 12 00:01:00.700 --> 00:01:08.090 Sadia Afrin Purba: hopefully what she talks about today will begin her journey towards a dissertation in a not-too-long time, so… 13 00:01:08.460 --> 00:01:12.210 Sadia Afrin Purba: The format is the first hour is public, open to the public. 14 00:01:12.570 --> 00:01:15.460 Sadia Afrin Purba: People are encouraged to ask questions as they go along. 15 00:01:15.600 --> 00:01:28.329 Sadia Afrin Purba: And then we will excuse the public, everyone about the committee, and the committee will spend some time asking her questions on her materials. And then the committee will defer and… 16 00:01:28.450 --> 00:01:30.890 Sadia Afrin Purba: Be informed of the decision shortly thereafter. 17 00:01:31.120 --> 00:01:33.060 Sadia Afrin Purba: So, Sally, you're welcome, and good luck. 18 00:01:33.200 --> 00:01:35.170 Sadia Afrin Purba: Thank you for the introduction. 19 00:01:35.370 --> 00:01:43.089 Sadia Afrin Purba: Hello, my name is Shadyafing Purva. I am here to talk about the quantum correlation enabling quantum advantage in machine learning. 20 00:01:43.220 --> 00:01:52.109 Sadia Afrin Purba: So I hope by… after this presentation, everyone has some idea about quantum computing, and why it is very important for machine learning. 21 00:01:52.480 --> 00:02:01.000 Sadia Afrin Purba: Okay, so my outline is we have many definitions, like quantum advantage, quantum correlation, so I will define those first. 22 00:02:01.180 --> 00:02:08.229 Sadia Afrin Purba: Then, I'll talk about my motivation, why the quantum correlation is very important for machine learning tasks. 23 00:02:08.570 --> 00:02:14.890 Sadia Afrin Purba: Then I'll talk about some fundamental quantum computing and quantum information, related topic. 24 00:02:15.140 --> 00:02:28.330 Sadia Afrin Purba: Then the main issue, the main topic, I'll… I have selected three, papers which actually, prove, quantum advantage in machine learning field, so I'll discuss their contribution. 25 00:02:28.380 --> 00:02:38.349 Sadia Afrin Purba: And based on those papers, I have created… I come up some roadmap that we need to focus on for our future research in quantum machine learning. 26 00:02:38.920 --> 00:02:39.830 Sadia Afrin Purba: Sure. 27 00:02:40.680 --> 00:02:43.779 Sadia Afrin Purba: What is actually… what is quantum, actually? I mean… 28 00:02:43.920 --> 00:02:56.409 Sadia Afrin Purba: Quantum deal with subatomic particles. So, when we are talking about quantum mechanics, we are talking about the physics that are, that are in the subatomic 29 00:02:56.440 --> 00:03:08.559 Sadia Afrin Purba: Particles, for example, electron and photon, and many physicians and computer scientists describe the quantum mechanics as a generalization of classical probability theory. 30 00:03:09.130 --> 00:03:18.290 Sadia Afrin Purba: And in quantum mechanics, we have two important probiotics, and these are very important. One is superposition, and another is entanglement. 31 00:03:19.010 --> 00:03:24.540 Sadia Afrin Purba: And quantum computing is a type of computation that uses these two 32 00:03:24.680 --> 00:03:32.460 Sadia Afrin Purba: Basically, these two, properties to process information in a way that the classical computer cannot do. 33 00:03:33.550 --> 00:03:43.749 Sadia Afrin Purba: And by quantum advantage, I meant that we have a goal that we want to prove that a programmable quantum computer can solve a problem. 34 00:03:43.850 --> 00:03:53.740 Sadia Afrin Purba: That no classical computer can solve in any feasible amount of time, irrespective of the usefulness of the problem. 35 00:03:53.820 --> 00:04:07.679 Sadia Afrin Purba: This particular term was coined by Dr. Preskill in 2011, but the concept of having a quantum computer for simulating our nature is from early 80s. 36 00:04:08.380 --> 00:04:22.020 Sadia Afrin Purba: So, there are… there are three examples. For example, in 2019, Google, showed the, random shark gate sampling that would take only 200 seconds. 37 00:04:22.930 --> 00:04:31.610 Sadia Afrin Purba: In quantum computer, but it would take 10,000 years for a, supercomputer, so that is a huge speedup. 38 00:04:31.730 --> 00:04:35.209 Sadia Afrin Purba: Another is a very famous algorithm, Google's algorithm. 39 00:04:35.340 --> 00:04:51.209 Sadia Afrin Purba: It provides quadratic speedup, if we run the unstructured search in a quantum computer. And then this is a famous Shores algorithm, which is… which deals with integer factorization. 40 00:04:51.250 --> 00:05:09.310 Sadia Afrin Purba: And if it is… so far, it is, theoretical, but when we have a… when we have a quantum computer that can perform the source algorithm, the internet is broken, the RESA, increasingly is broken, and that particular day is called Q Day. 41 00:05:09.530 --> 00:05:22.369 Sadia Afrin Purba: Now, when we are talking about speedup, a natural question is, why do we need speedup? I mean, we have many cluster high-performance computers, why do we need that? 42 00:05:22.530 --> 00:05:33.750 Sadia Afrin Purba: Now, the concept is, for example, there is one algorithm that solves a problem, and the complexity is O of, O of root… O . 43 00:05:34.040 --> 00:05:42.790 Sadia Afrin Purba: And no matter how much, how many classical computers are we running, the internal complexity does not change. 44 00:05:43.050 --> 00:05:44.010 Sadia Afrin Purba: But. 45 00:05:44.290 --> 00:05:53.369 Sadia Afrin Purba: Quantum computing, and with help of superposition and entanglement, we have the power to reduce the complexity. So it is a… 46 00:05:53.470 --> 00:05:58.700 Sadia Afrin Purba: It is a different kind of computation, so it is not like apple-to-apple comparison. 47 00:05:59.960 --> 00:06:02.720 Sadia Afrin Purba: So, let me ask a question about the second… 48 00:06:03.800 --> 00:06:06.960 Sadia Afrin Purba: There's the first and third, somewhat familiar with the second one. 49 00:06:07.450 --> 00:06:09.040 Sadia Afrin Purba: the program's algorithm. Yeah. 50 00:06:09.710 --> 00:06:14.359 Sadia Afrin Purba: Can you say a little bit more about the speedup, is this a theoretical speedup? Is this… 51 00:06:14.790 --> 00:06:22.719 Sadia Afrin Purba: And speed up the… It's demonstrated in the… Things are… Exactly under what conditions. 52 00:06:23.510 --> 00:06:42.980 Sadia Afrin Purba: what's disproven to get this kind of a speedup? Is it parallelism? Is it what… what's giving it this, rather dramatic speed up in search? It specifically used superposition property very well. So, it simultaneously, starts the… 53 00:06:43.260 --> 00:06:47.100 Sadia Afrin Purba: Search the database with a specific key. 54 00:06:47.250 --> 00:07:06.789 Sadia Afrin Purba: and simultaneously get the result within a… I mean, with a root over n, in time complexity. It is first… first of all, it is a theoretical bound, but, we have, there are experimentation in current hardware that it is true. We can get, 55 00:07:06.950 --> 00:07:15.880 Sadia Afrin Purba: with this ratio, with a limited number of… for example, I think it, the experimentation was shown with 20 qubits. 56 00:07:17.120 --> 00:07:18.570 Sadia Afrin Purba: And this is just an across… 57 00:07:19.620 --> 00:07:37.389 Sadia Afrin Purba: Yeah. Unshorted. No, no, nothing. Just a database search. I mean, unsorted, you have a key, you have an unsorted data set, you are… you want to search the data. 58 00:07:37.500 --> 00:07:38.190 Sadia Afrin Purba: Make reasoning. 59 00:07:39.130 --> 00:07:41.729 Sadia Afrin Purba: I, the superposition. 60 00:07:42.360 --> 00:07:44.649 Sadia Afrin Purba: Because… So it's being able to do… 61 00:07:45.760 --> 00:07:49.600 Sadia Afrin Purba: It's a form of parallelism. Yeah, yeah. We do multiple things. 62 00:07:49.800 --> 00:07:51.919 Sadia Afrin Purba: That's one. Yeah. 63 00:07:56.450 --> 00:07:58.190 Sadia Afrin Purba: Now, 64 00:07:58.800 --> 00:08:14.940 Sadia Afrin Purba: Now, classical correlation, we all know, it is, it describes, how two random variables are linearly, related, but our classical measures obey by classical, probability theory and local realism theory. 65 00:08:14.940 --> 00:08:20.300 Sadia Afrin Purba: What is local realism theory in classical world, before calculating the 66 00:08:20.340 --> 00:08:28.220 Sadia Afrin Purba: probability, we have all the information about the random variable. I mean, we can… we can know the prior information, right? 67 00:08:28.930 --> 00:08:48.479 Sadia Afrin Purba: the nearby events influence each other. For example, we are dealing with a coin. If we are dealing with coin, the nearby coin will influence the probability. It is not like one coin is here, another coin is here. 68 00:08:48.660 --> 00:08:57.880 Sadia Afrin Purba: There is no communication, so how do you measure the, measure the probability? In classical world, there is no way. 69 00:08:57.970 --> 00:09:08.900 Sadia Afrin Purba: But in quantum correlation, it is a general term of quantum mechanics that refers to non-classical statistical relationship between two or more, particles. 70 00:09:09.220 --> 00:09:18.829 Sadia Afrin Purba: For example, we have, like, two quantum dice, and they are very far away. One is here, another one is probably in another planet. 71 00:09:18.990 --> 00:09:21.429 Sadia Afrin Purba: And if I dice, the… 72 00:09:21.700 --> 00:09:30.839 Sadia Afrin Purba: if I roll the dice, it always… the both of the dice always give the… give the same output. So that is the entanglement. 73 00:09:31.420 --> 00:09:38.879 Sadia Afrin Purba: So the non-local pro… non-local feature is very important here in quantum correlation. 74 00:09:39.290 --> 00:09:53.329 Sadia Afrin Purba: And the strongest form is quantum entanglement. There is also a general type of weaker quantum correlation, which is called, quantum disc code. And my presentation is focused on this quantum co-entanglement and quantum disk code. 75 00:09:54.580 --> 00:10:05.449 Sadia Afrin Purba: So, what is the motivation? In machine learning tasks, our main goal is to model a distribution. Given X, we want to model the distribution of Y. 76 00:10:05.660 --> 00:10:07.670 Sadia Afrin Purba: And based on the training data. 77 00:10:08.280 --> 00:10:13.869 Sadia Afrin Purba: And we use some kind of parameter and loss function to learn that. 78 00:10:14.520 --> 00:10:19.560 Sadia Afrin Purba: Now, classical models rely on the joint and conditional probability. 79 00:10:19.750 --> 00:10:23.490 Sadia Afrin Purba: Which is only capture the statistical dependency. 80 00:10:24.070 --> 00:10:33.479 Sadia Afrin Purba: We can, no, not… traditionally, we cannot capture the physical relationship between the variable with, in the classical model. 81 00:10:33.820 --> 00:10:41.690 Sadia Afrin Purba: And correlation are local, meaning information is shared through the explicit data feature or parameter. 82 00:10:41.960 --> 00:10:51.489 Sadia Afrin Purba: And it is also constrained by cars of dimensity. For example, if we have more data, we need to model more 83 00:10:51.700 --> 00:10:56.139 Sadia Afrin Purba: More, hype, more hidden layer. 84 00:10:57.460 --> 00:11:05.880 Sadia Afrin Purba: So, my question… my motivation is, can non-local quantum correlation provide a technological or quantum advantage for machine learning tasks? 85 00:11:06.100 --> 00:11:09.250 Sadia Afrin Purba: So, do you have any questions at this point? 86 00:11:09.680 --> 00:11:10.670 Sadia Afrin Purba: Your definition? 87 00:11:11.330 --> 00:11:14.969 Sadia Afrin Purba: learning. Would you say that's a generative model or a standard model? 88 00:11:15.630 --> 00:11:21.320 Sadia Afrin Purba: For this particular case, I choose a… supervise. 89 00:11:21.480 --> 00:11:24.330 Sadia Afrin Purba: modern, I mean… That wasn't the true. 90 00:11:24.970 --> 00:11:31.220 Sadia Afrin Purba: Generative or disternative? Dominative. Dominative. That equation, you believe, describes a determinative model. 91 00:11:33.810 --> 00:11:40.150 Sadia Afrin Purba: No, I didn't put the, What? Joint probability. 92 00:11:43.060 --> 00:11:51.259 Sadia Afrin Purba: Well, I think it's a narrow… it's a narrow definition of machine learning. Yeah, yeah. Because this is more what you expect of a generic model that you're able to 93 00:11:52.190 --> 00:12:04.080 Sadia Afrin Purba: understand the statistics. We don't really… in classification, you don't really care about the statistics, you just care about your ability to learn the mapping. Yeah, but how do you map that? You need… 94 00:12:04.160 --> 00:12:12.310 Sadia Afrin Purba: you need the similarity or some kind of correlation, the dot product, that is. Right, but you don't really care if it… 95 00:12:12.600 --> 00:12:15.900 Sadia Afrin Purba: If you can directly model… you can accurately model the 96 00:12:16.010 --> 00:12:22.589 Sadia Afrin Purba: The conditional distributions and things, you just care about the ability to do the class assignments. 97 00:12:24.140 --> 00:12:25.020 Sadia Afrin Purba: Bingham. 98 00:12:25.450 --> 00:12:32.509 Sadia Afrin Purba: Okay. Yeah, I should have put a very big, general definition, yeah. 99 00:12:33.180 --> 00:12:33.900 Sadia Afrin Purba: Good. 100 00:12:35.510 --> 00:12:50.610 Sadia Afrin Purba: Now, in quantum ML, the quantum advantage is the ability of quantum systems to perform a learning task. As we are talking about machine learning, we focus on learning tasks faster, with fewer ratios, and with greater or greater accuracy. Any of them. 101 00:12:50.650 --> 00:12:55.740 Sadia Afrin Purba: If we achieve any of them, we can say that, yeah, we achieve quantum advantage. 102 00:12:56.310 --> 00:12:57.230 Sadia Afrin Purba: And… 103 00:12:58.020 --> 00:13:06.840 Sadia Afrin Purba: I am… I am arguing that quantum correlation can, learn information beyond the training distribution, because it has some 104 00:13:07.170 --> 00:13:20.919 Sadia Afrin Purba: capability to learn the global information. It does not only, process, variables locally, it has the idea of… the, idea of the system global… in a global sense. And I think the… 105 00:13:21.110 --> 00:13:25.339 Sadia Afrin Purba: In classical world, we also only deal with the local information. 106 00:13:26.560 --> 00:13:33.100 Sadia Afrin Purba: And also, the phase correlation allow contrastive interface, which, 107 00:13:33.240 --> 00:13:36.630 Sadia Afrin Purba: Which actually boosts your learning and efficiency. 108 00:13:36.740 --> 00:13:42.830 Sadia Afrin Purba: And, cancel out the wrong answer, and boost the right answer. 109 00:13:43.080 --> 00:13:50.009 Sadia Afrin Purba: And also the entanglement-based correlation It is a very, efficient. 110 00:13:50.270 --> 00:14:01.719 Sadia Afrin Purba: information transfer, transfer system. I mean, if two systems are correlated, if I measure one system, I… I don't need to measure the other system. 111 00:14:02.010 --> 00:14:08.499 Sadia Afrin Purba: So, the… The information transformation or the communication bottleneck can be… 112 00:14:08.600 --> 00:14:13.430 Sadia Afrin Purba: Can be reduced by using entanglement correlation. 113 00:14:16.880 --> 00:14:21.440 Sadia Afrin Purba: What do you mean between the difference between global dependencies and global dependencies? 114 00:14:21.660 --> 00:14:22.919 Sadia Afrin Purba: It's a little bit lost. 115 00:14:23.160 --> 00:14:28.540 Sadia Afrin Purba: Okay, for example, Like, global… 116 00:14:28.580 --> 00:14:32.010 Sadia Afrin Purba: For example, we want to, 117 00:14:32.040 --> 00:14:42.479 Sadia Afrin Purba: we want to understand that if it is a… if… we want to understand an image classification task, okay? So, we only understand 118 00:14:42.480 --> 00:14:53.579 Sadia Afrin Purba: We can only model the image classification task in terms of the training data. We are… the model can never inherit knowledge. 119 00:14:53.670 --> 00:14:57.420 Sadia Afrin Purba: From beyond the training distribution. 120 00:14:59.910 --> 00:15:02.390 Sadia Afrin Purba: Okay, so that is a local part. 121 00:15:02.700 --> 00:15:15.070 Sadia Afrin Purba: Now, the quantum is the… Quantum, quantum entanglement deal with, subsystem, okay? Now, it has… 122 00:15:15.210 --> 00:15:24.800 Sadia Afrin Purba: Some, it has to… Not all knowledge, it has the whole perspective, a global perspective from the 123 00:15:24.970 --> 00:15:32.119 Sadia Afrin Purba: each subs, system. So that is the global, global perspective. And from that global perspective. 124 00:15:32.270 --> 00:15:40.979 Sadia Afrin Purba: I mean, it is in a third-party perspective, so it can actually, extract the feature. 125 00:15:41.420 --> 00:15:44.169 Sadia Afrin Purba: Beyond the training dataset, whatever. 126 00:15:44.720 --> 00:15:45.740 Sadia Afrin Purba: Providing it. 127 00:15:46.030 --> 00:15:52.629 Sadia Afrin Purba: I'm getting work today. Okay. Can you go back to your example of the image processing? Yeah. 128 00:15:53.240 --> 00:15:57.219 Sadia Afrin Purba: what a global dependency would be. 129 00:15:57.910 --> 00:16:07.150 Sadia Afrin Purba: Image processing, or whatever, like, Okay. For example, you have, trained your… 130 00:16:07.580 --> 00:16:17.269 Sadia Afrin Purba: train your image classifier with a certain breed of dog, only a certain breed of dog and certain breed of cat, okay? 131 00:16:17.500 --> 00:16:18.639 Sadia Afrin Purba: Nothing else. 132 00:16:18.740 --> 00:16:27.299 Sadia Afrin Purba: Okay. So, do you think if you give him a very odd-looking, but that is also a dog. 133 00:16:27.680 --> 00:16:31.120 Sadia Afrin Purba: to the email… to the model, it will classify. 134 00:16:31.910 --> 00:16:34.050 Sadia Afrin Purba: Depends on how similar that is. 135 00:16:34.230 --> 00:16:54.119 Sadia Afrin Purba: Oh, no, it is not very similar, but it is a dog. Yeah. So, in quantum world, suppose all the dog, the, all the breeds, they are dog, right? They are entangled. So, when I learn, when I, make my model learn one kind of dog. 136 00:16:54.660 --> 00:17:13.180 Sadia Afrin Purba: due to the entanglement property, I know this is a dog. I mean, the property… I have learned that it's a dog. So, due to the correlation, the quantum correlation, not the classical one, the quantum correlation, it is theoretically possible that, yeah, I can… 137 00:17:13.339 --> 00:17:18.520 Sadia Afrin Purba: learn about, or I can understand, I can recognize any kind of dog in the world. 138 00:17:18.650 --> 00:17:24.919 Sadia Afrin Purba: Is that because the pixels within the image are somehow quantitativeled? Yeah, 139 00:17:25.690 --> 00:17:30.739 Sadia Afrin Purba: Not on… Physically, I'm the same, I don't get it physically, how it works. I… 140 00:17:31.010 --> 00:17:36.520 Sadia Afrin Purba: I know a little bit of quantum from many years ago, but I still don't get… So what… 141 00:17:36.910 --> 00:17:46.100 Sadia Afrin Purba: A little more specific. You have a stack of pictures of dogs. Yeah. Right? A stack of pictures of cats. Yeah, and these are quant… entangled data. 142 00:17:46.390 --> 00:17:56.799 Sadia Afrin Purba: How can you say that? What does that mean? I, I have some, I have some classical feature. I run the classical feature to a entangled circuit. 143 00:17:56.950 --> 00:17:59.579 Sadia Afrin Purba: I get some quantum state, right? 144 00:17:59.740 --> 00:18:03.750 Sadia Afrin Purba: So that is a… entangle dataset. 145 00:18:04.000 --> 00:18:06.020 Sadia Afrin Purba: So you have to prepare the entanglement. 146 00:18:06.380 --> 00:18:08.830 Sadia Afrin Purba: Yeah. Which involves adding information. 147 00:18:09.520 --> 00:18:21.730 Sadia Afrin Purba: No, it involves representing information. Yeah, not adding information, representing the information in a different dimension. Okay. 148 00:18:22.090 --> 00:18:26.520 Sadia Afrin Purba: Would that take as much time as the time you saved by being able to do the consultation faster? 149 00:18:27.230 --> 00:18:34.220 Sadia Afrin Purba: Okay, I'm… I mean, the… could you do the same thing in regular… 150 00:18:34.540 --> 00:18:39.050 Sadia Afrin Purba: non-quantum machine learning. I mean, you could say, you could teach it, this is what 151 00:18:39.170 --> 00:18:54.600 Sadia Afrin Purba: breed number one looks like, and this is what breed number… How many data points do you need? Meaning… You're talking about entangling all these data points. Yeah, if I use entanglement, okay, it is possible that I need only one document to classify all the documents. 152 00:18:54.920 --> 00:18:56.790 Sadia Afrin Purba: It is theoretically possible. 153 00:18:57.060 --> 00:19:05.520 Sadia Afrin Purba: Okay, but in classically, if I want to classify all breed of all breed of dog. 154 00:19:05.800 --> 00:19:08.529 Sadia Afrin Purba: I need probably trillions of images. 155 00:19:08.650 --> 00:19:10.660 Sadia Afrin Purba: So they are… 156 00:19:14.370 --> 00:19:25.320 Sadia Afrin Purba: I have to create the entangled… yeah, I have to create the entangled dataset. That is a very big research area, how to do this. 157 00:19:25.860 --> 00:19:27.079 Sadia Afrin Purba: We gotta talk about that. 158 00:19:27.790 --> 00:19:31.339 Sadia Afrin Purba: Let's probably going concrete about this at that time. 159 00:19:31.980 --> 00:19:34.329 Sadia Afrin Purba: The data itself could be a table. 160 00:19:34.500 --> 00:19:38.590 Sadia Afrin Purba: We've talked about data sets that are clearly from entangled data, like 161 00:19:38.930 --> 00:19:44.859 Sadia Afrin Purba: One of the sets I've been trying to get her interested in is traffic flow on the Google Express. Yeah. 162 00:19:45.310 --> 00:19:46.980 Sadia Afrin Purba: Because you wouldn't believe that… 163 00:19:47.140 --> 00:19:51.890 Sadia Afrin Purba: The way people behave is a handled in some way, and it's not a resentment like the raw data. 164 00:19:52.080 --> 00:20:01.549 Sadia Afrin Purba: But now you're talking about actually creating a model of entanglement, like artificial entanglement. Yeah, because we have to… we have to… 165 00:20:02.280 --> 00:20:06.750 Sadia Afrin Purba: You, why do you think it is a limitation? We have to… 166 00:20:06.950 --> 00:20:13.030 Sadia Afrin Purba: Whatever sort of differentiates between what they should really call it something like artificial intelligence. 167 00:20:13.810 --> 00:20:20.340 Sadia Afrin Purba: No, I don't… okay, probably, probably. And also, how do you model the… 168 00:20:20.940 --> 00:20:28.839 Sadia Afrin Purba: I mean, we have all… we have, I think, all, experienced some kind of spooky event, like. 169 00:20:29.190 --> 00:20:31.210 Sadia Afrin Purba: I… out of nowhere. 170 00:20:31.480 --> 00:20:44.300 Sadia Afrin Purba: intuition comes, and yeah, you feel like you should not go there, you didn't go there, and something bad happened. Okay. Similar… similar kind of the mother-child telepathic. 171 00:20:44.370 --> 00:20:54.579 Sadia Afrin Purba: This is real. The mother tell everybody, it happened with me many times, with my mother. How do you model such thing with the classical… any kind of learning mission? 172 00:20:55.660 --> 00:20:57.950 Sadia Afrin Purba: So they are beyond some… 173 00:20:58.250 --> 00:21:04.750 Sadia Afrin Purba: event that cannot model with classical machine learning or classical algorithm. That is not possible. 174 00:21:06.330 --> 00:21:09.420 Sadia Afrin Purba: But that's possible to really support them? Yeah. 175 00:21:09.960 --> 00:21:25.020 Sadia Afrin Purba: Yeah, I mean, I am not saying it's possible, I am saying that there are reviews potential that if we exploit the entanglement or quantum correlation, this property in our modeling, in our mathematical 176 00:21:25.290 --> 00:21:30.019 Sadia Afrin Purba: Modeling, then we can extract more information. 177 00:21:30.180 --> 00:21:33.169 Sadia Afrin Purba: That we cannot by doing classical algorithm. 178 00:21:34.880 --> 00:21:43.660 Sadia Afrin Purba: I mean, physicists in the lab have created particles that entangled things on purpose. Yeah. And then collapsed it by observing it. Yeah. 179 00:21:44.160 --> 00:21:52.150 Sadia Afrin Purba: Yeah, so that… with that concept, we want to simulate that mathematically in our machine learning task. 180 00:21:53.780 --> 00:21:55.849 Sadia Afrin Purba: So what do you mean by coherent influence? 181 00:21:56.050 --> 00:21:59.400 Sadia Afrin Purba: That's… Okay, cool. 182 00:22:00.220 --> 00:22:09.270 Sadia Afrin Purba: For example, the… image classification problem. Quantum image classification problem. I have a quantum… Yeah, I have… 183 00:22:10.240 --> 00:22:13.860 Sadia Afrin Purba: Okay? But I have a quantum in it, okay? 184 00:22:14.100 --> 00:22:16.699 Sadia Afrin Purba: I have a quantum document. 185 00:22:17.610 --> 00:22:22.029 Sadia Afrin Purba: I have prepared the quantum entangle somehow, so I have a quantum document. 186 00:22:22.980 --> 00:22:24.969 Sadia Afrin Purba: And I learned that. 187 00:22:25.240 --> 00:22:25.950 Sadia Afrin Purba: King. 188 00:22:26.130 --> 00:22:27.060 Sadia Afrin Purba: No. 189 00:22:27.670 --> 00:22:36.250 Sadia Afrin Purba: I can, I mean, now I get a, new quantum document with a new breed, okay? 190 00:22:36.690 --> 00:22:37.460 Sadia Afrin Purba: No. 191 00:22:37.970 --> 00:22:47.560 Sadia Afrin Purba: I mean, I can instantly know that that is a dog. This is an instant, information transfer. I don't need a communication channel. 192 00:22:48.190 --> 00:23:00.449 Sadia Afrin Purba: I mean, yeah, in practice, if we want to build a quantum internet, we need a classical channel and all that stuff, but let's just not focus on that. 193 00:23:00.670 --> 00:23:14.549 Sadia Afrin Purba: without any communication, without any measurement, I can know that, yeah, there is another image that is, a dog. So that is the coherent… and that is a right information, not a wrong information. 194 00:23:14.610 --> 00:23:27.130 Sadia Afrin Purba: In classical world, if you want to communicate that, you need some kind of channel, or you need some kind of bit transformation, right? In quantum, there is no bit transformation. 195 00:23:29.270 --> 00:23:29.960 Sadia Afrin Purba: Okay. 196 00:23:34.580 --> 00:23:42.789 Sadia Afrin Purba: Okay, then I have selected these, three papers. The first one actually, described the entangled base. 197 00:23:43.160 --> 00:23:45.539 Sadia Afrin Purba: Quantum advantage for them. 198 00:23:45.690 --> 00:24:03.349 Sadia Afrin Purba: learning task, and another one is a quantum freelance theory. It is a theoretical paper, and this paper actually theoretically proved that if we have entangled data, then only one, training sample is enough to learn the whole system. 199 00:24:03.690 --> 00:24:06.380 Sadia Afrin Purba: This is the theoretical proof of this paper. 200 00:24:06.510 --> 00:24:14.300 Sadia Afrin Purba: And another is, discourse is also a quantum correlation, so it is a discourse code-based advantage. 201 00:24:15.220 --> 00:24:30.450 Sadia Afrin Purba: Okay, then it is just timeline. I want to, give an impression that we are very early in a quantum computing, and very early in a quantum computing with machine learning, so we should be very open-minded and 202 00:24:31.190 --> 00:24:37.369 Sadia Afrin Purba: You know… optimistic. At least I am very optimistic. So, these are the… 203 00:24:37.470 --> 00:24:51.190 Sadia Afrin Purba: mainly focused on entanglement. It's actually acquired by Strodinger in 1935. Einstein didn't believe it, but in 1964, a theory came, Bell's theorem, which actually challenged Einstein's view. 204 00:24:51.310 --> 00:25:02.570 Sadia Afrin Purba: And in 1972, it was proven that, yeah, entanglement is a true thing. It is not something theoretical. We can create entanglement, so Einstein was wrong in this particular case. 205 00:25:03.110 --> 00:25:19.230 Sadia Afrin Purba: And in 2009, there is the first algorithm, HHL algorithm. I don't know about very deeply about this algorithm, but this particular algorithm was invented to solve the linear system. 206 00:25:19.900 --> 00:25:23.119 Sadia Afrin Purba: And what did Einstein say about quantum mechanics? 207 00:25:23.420 --> 00:25:36.229 Sadia Afrin Purba: There is two… yeah, there is two famous quotes that God does not play with dice, and another quote is, I would like to believe that the moon is there when I am not looking at it. 208 00:25:36.810 --> 00:25:47.199 Sadia Afrin Purba: Because in quantum, mechanics, there is an interpretation that when I… there is an object, only when I see it, there is no real existence of this object. 209 00:25:47.490 --> 00:25:50.459 Sadia Afrin Purba: But then we also say, if you think you understand what I'm saying? 210 00:25:52.120 --> 00:25:55.540 Sadia Afrin Purba: I don't understand… paramilia. 211 00:25:56.240 --> 00:25:57.030 Sadia Afrin Purba: women. 212 00:25:57.940 --> 00:26:03.889 Sadia Afrin Purba: Yeah, I don't fully understand quantum mechanics, but I understand enough to appreciate it. 213 00:26:06.710 --> 00:26:13.920 Sadia Afrin Purba: Okay, now, the… But we're in this… Nice area. 214 00:26:14.170 --> 00:26:23.809 Sadia Afrin Purba: Like, it does things for you. Yeah. Okay, so even though we're only in a disk, it's good enough to do stuff with them now. Tisha, currently we… 215 00:26:24.280 --> 00:26:36.470 Sadia Afrin Purba: Yeah, currently we have a thousand… mostly thousand qubits, and also those are very noisy, so we have to create error-corrected algorithm and all that stuff. 216 00:26:36.860 --> 00:26:45.350 Sadia Afrin Purba: Okay, so we are here right now, and probably within 10 years, we will have a full fault-tolerant quantum computer. 217 00:26:46.450 --> 00:26:51.540 Sadia Afrin Purba: Okay, now some fundamental. Now, I, really like to, 218 00:26:51.680 --> 00:27:00.879 Sadia Afrin Purba: Now, in the fundamental, I am going to talk about the qubits. Now, the best way to represent qubit is, like, electron and its nucleus. 219 00:27:01.470 --> 00:27:10.889 Sadia Afrin Purba: So, we are… I always think that electrons stay either CAT0 or CAT1, either ground state or excited state. But in… in practice. 220 00:27:11.050 --> 00:27:23.529 Sadia Afrin Purba: it cannot make up its mind. I mean, it is a probability… a cloud of probability. For example, the red portion means that, yeah, here is a high probability that we can find the electron in here. 221 00:27:23.550 --> 00:27:36.930 Sadia Afrin Purba: And the blue one means that, yeah, there are low probability. So it is not discrete. There are some kind of probability and, cloud of probabilities going on. So how can we 222 00:27:37.010 --> 00:27:46.150 Sadia Afrin Purba: represent the… information. We can represent that CAT0 and CAT1. This is the quantum bit. 223 00:27:46.970 --> 00:27:59.830 Sadia Afrin Purba: With alpha and beta. These are probability amplitudes. It, it is telling us that, if we, it is telling us that, this side, the quantum state, in, 224 00:28:00.450 --> 00:28:09.410 Sadia Afrin Purba: Could be in a ground state with alpha probability, alpha square probability, or in a 225 00:28:10.130 --> 00:28:13.609 Sadia Afrin Purba: Excited state in a beta square probability. 226 00:28:14.000 --> 00:28:29.670 Sadia Afrin Purba: This alpha and beta itself, probability amplitude, not probability, because it could be minus 1, it could be a complex number. So, from there, we can get probability, but these are a probability amplitude, so there's a slight difference. Any question? 227 00:28:30.020 --> 00:28:34.200 Sadia Afrin Purba: Do you know the paper that that electron bill paper indicates on that? 228 00:28:34.390 --> 00:28:37.900 Sadia Afrin Purba: Yeah, it is from probably 2003's paper. 229 00:28:38.120 --> 00:28:39.410 Sadia Afrin Purba: Yeah. 230 00:28:39.550 --> 00:28:41.529 Sadia Afrin Purba: I don't remember the name. 231 00:28:41.670 --> 00:28:44.150 Sadia Afrin Purba: Oh, okay. 232 00:28:44.650 --> 00:28:47.010 Sadia Afrin Purba: Okay. 233 00:28:47.200 --> 00:28:52.390 Sadia Afrin Purba: Yeah, I got the image from that. 234 00:28:52.580 --> 00:28:53.400 Sadia Afrin Purba: Yeah. 235 00:28:53.680 --> 00:28:55.499 Sadia Afrin Purba: Yeah, that is cool. 236 00:28:55.690 --> 00:29:10.550 Sadia Afrin Purba: Okay, so the difference is classical bit can be 0 or 1, right? We can never get any kind of probability in beating, but in quantum bit, there is a probability that we can get here. 237 00:29:10.660 --> 00:29:12.750 Sadia Afrin Purba: Which is a superposition. 238 00:29:12.910 --> 00:29:19.060 Sadia Afrin Purba: And if we measure that, we will have a, classical probability. 239 00:29:19.560 --> 00:29:23.459 Sadia Afrin Purba: So, I would like to view in a, vector form. 240 00:29:23.600 --> 00:29:25.089 Sadia Afrin Purba: CAT0 means 241 00:29:25.210 --> 00:29:41.590 Sadia Afrin Purba: 1, 0. I like to say is that Cat0 means the probability of finding the qubit in a ground state, or upspring, or any kind of physical property is 100%, because the alpha for this position, the zero position, is 100%. 242 00:29:41.590 --> 00:29:49.210 Sadia Afrin Purba: Similarly, CAT1 means that it is in a 100% probability that I will find the electron in a ground state. 243 00:29:49.990 --> 00:30:01.169 Sadia Afrin Purba: And more generally, we can write in this format, and by doing the square, magnitude square, we will get a classical probability, and the classical probability has to be 1. 244 00:30:02.560 --> 00:30:03.450 Sadia Afrin Purba: Sorry. 245 00:30:04.260 --> 00:30:14.249 Sadia Afrin Purba: Okay, now, we can also represent this in a blot sphere. It is a sphere. This is our zero cat, this is… this is our one cat, and there are… 246 00:30:14.550 --> 00:30:18.340 Sadia Afrin Purba: Many states where the, electron can be. 247 00:30:18.860 --> 00:30:28.379 Sadia Afrin Purba: And this is our quantum circuit, it is also the quantum algorithm or quantum model, whatever it is. Now, this is nothing but linear sequence of 248 00:30:28.630 --> 00:30:32.669 Sadia Afrin Purba: linear transformation of quantum gate. Quantum gate is also, like. 249 00:30:32.940 --> 00:30:38.980 Sadia Afrin Purba: Some metrics, some specific metrics that are helpful to do the linear transformation. 250 00:30:40.040 --> 00:30:43.160 Sadia Afrin Purba: And the quantum gate has to be unitary. 251 00:30:43.610 --> 00:30:52.460 Sadia Afrin Purba: It has to be unitary, and Hadamat is a very famous git to create the create the… superposition. 252 00:30:52.770 --> 00:31:05.830 Sadia Afrin Purba: And Poly XYZ, they also rotate the computational basis in a particular angle and all that stuff. And for, creating the entanglement, we need the CNOT gate. 253 00:31:06.470 --> 00:31:24.570 Sadia Afrin Purba: And why we need to… why the quantum gate has to be unitary? Because we want to reserve the probability. And unitary has some very cool, feature from linear algebra, so we need the gate to be unitary. 254 00:31:24.900 --> 00:31:33.589 Sadia Afrin Purba: And how can we do the measurement? By convention, we always use a jet basis to do the measurement. When we measure. 255 00:31:34.200 --> 00:31:39.770 Sadia Afrin Purba: The electron makeup is mine, and it can be only 0 or 1. 256 00:31:40.560 --> 00:31:50.239 Sadia Afrin Purba: So, it is just, more in detail that how it is happening. I mean, if I apply Hadamatt gate to the zero state. 257 00:31:50.240 --> 00:32:06.840 Sadia Afrin Purba: Then simply metric multiplication, and we will get 1 by root over 2, 1 by root over 2, which is 50%, and that is why it is the… it is very famous to say that a qubit can be in a state of both 0 and 1 at the both… at the same thing. 258 00:32:07.240 --> 00:32:08.910 Sadia Afrin Purba: This is the superposition. 259 00:32:10.950 --> 00:32:15.499 Sadia Afrin Purba: And, this is a very, simple start kit. 260 00:32:15.830 --> 00:32:28.449 Sadia Afrin Purba: And if we calculate the probability, we can see that it is 50-50%. After measurement, the electron makeup, it mines and, collapses through one, one, one basis. 261 00:32:28.620 --> 00:32:29.630 Sadia Afrin Purba: Any question? 262 00:32:31.310 --> 00:32:44.029 Sadia Afrin Purba: cool, but you need it? Because I guess it encapsulates the… For the single qubit, I think the visualization. For the visualization, we need that. Yeah. 263 00:32:45.170 --> 00:32:47.410 Sadia Afrin Purba: Yeah, we can do that. 264 00:32:48.880 --> 00:32:50.059 Sadia Afrin Purba: There's a maintenance. 265 00:32:54.840 --> 00:32:57.710 Sadia Afrin Purba: Yeah, it is 556 chance before the measurement. 266 00:32:58.040 --> 00:32:58.900 Sadia Afrin Purba: Okay. 267 00:32:59.650 --> 00:33:00.450 Sadia Afrin Purba: One. 268 00:33:01.390 --> 00:33:16.399 Sadia Afrin Purba: No, when I measure, I measure in a particular basis, J basis, and when I measure it, with… there is a form, measurement, postulate, I mean, we… there involve eigenvalue and all that stuff. 269 00:33:16.620 --> 00:33:30.949 Sadia Afrin Purba: And then it makes up your mind that, yeah, it is 50-50, so randomly choose zero, when it is 50-50. But suppose, suppose in this case, in the zero, there is a high probability. If I run 270 00:33:31.480 --> 00:33:41.060 Sadia Afrin Purba: many times, then it will always, collapse to zero, the highest state. For 50-50, it takes randomly. It shows randomly. 271 00:33:41.390 --> 00:33:43.090 Sadia Afrin Purba: And so what am I supposed to work. 272 00:33:43.840 --> 00:33:47.640 Sadia Afrin Purba: Yeah, yeah, yeah, because, probability has to be 1. 273 00:33:50.910 --> 00:33:54.789 Sadia Afrin Purba: But we could have just received even outcome 1 instead of outcome 0. Yeah. 274 00:33:56.640 --> 00:34:04.460 Sadia Afrin Purba: So, it is also another circuit, where, the zero… Zero probability is higher. 275 00:34:06.280 --> 00:34:12.940 Sadia Afrin Purba: And if I use 2 qubits, then we have equal, equal probability. 276 00:34:13.360 --> 00:34:22.350 Sadia Afrin Purba: But it is not fun, you know? If we use only superposition, we will always get the equal probability. It is not fun. 277 00:34:22.510 --> 00:34:29.299 Sadia Afrin Purba: And… The… for the… that particular thing, we need correlation. 278 00:34:30.020 --> 00:34:36.000 Sadia Afrin Purba: That is why the correlation is important, not only superposition. With superposition, we are not getting anywhere. 279 00:34:36.860 --> 00:34:46.509 Sadia Afrin Purba: So, we can create the correlation by CNOT gate. It is very simple, it is a control gate, and it is a target gate. We will change the target. 280 00:34:46.659 --> 00:34:52.680 Sadia Afrin Purba: We will flip the target if only the control bit is 1. This is a simple logic. 281 00:34:54.370 --> 00:34:57.020 Sadia Afrin Purba: By doing that, okay, previously. 282 00:34:57.650 --> 00:35:05.730 Sadia Afrin Purba: When we use only Hartamatin, rotation gate, even if we don't use the rotation gate, the probability will be equal. 283 00:35:06.040 --> 00:35:13.930 Sadia Afrin Purba: But, when we… When we use the CNOT gate, It is kind of, 284 00:35:14.790 --> 00:35:21.900 Sadia Afrin Purba: extra logic, right. So, this particular 0, 1, and 1, 0, this state can never basis. 285 00:35:22.070 --> 00:35:29.870 Sadia Afrin Purba: Right? So, we get zero probability, and we get 50-50% probability for 0, 0 and 11. 286 00:35:30.100 --> 00:35:38.989 Sadia Afrin Purba: So, correlation is important to filter out the equal probability thing, get some good information, and all that stuff. 287 00:35:41.550 --> 00:35:44.240 Sadia Afrin Purba: Then, I'll just skip this slide. 288 00:35:44.350 --> 00:35:49.340 Sadia Afrin Purba: I want to make the point that the definition of entanglement is 289 00:35:49.550 --> 00:36:02.810 Sadia Afrin Purba: Suppose we have a quantum state, and we cannot vector out that quantum state into an individual state. So, this was an example that proved that if we use C0, 290 00:36:02.990 --> 00:36:04.840 Sadia Afrin Purba: It has to be in Dengal. 291 00:36:06.420 --> 00:36:12.040 Sadia Afrin Purba: And… Then we have some other kind of, correlation, as I mentioned. 292 00:36:12.160 --> 00:36:21.009 Sadia Afrin Purba: that is quantum disc code. From my understanding, I think the quantum discord is a very generalized quantum correlation. For example, in this particular paper. 293 00:36:21.330 --> 00:36:31.410 Sadia Afrin Purba: they showed that up to this G position, mixing parameter J, up to this position, we get, entangled state. 294 00:36:31.780 --> 00:36:39.350 Sadia Afrin Purba: Below that position, we get some separable state, but we can get some correlation there also. 295 00:36:39.900 --> 00:36:45.600 Sadia Afrin Purba: Okay, so if we use 0, we have equal probability for all the BCs. 296 00:36:46.340 --> 00:36:51.129 Sadia Afrin Purba: If we increase the J value, the 0, 1, 297 00:36:51.670 --> 00:36:56.819 Sadia Afrin Purba: And one zero. This correlation, increased, and other decreased. 298 00:36:58.430 --> 00:36:59.510 Sadia Afrin Purba: Which one? 299 00:36:59.670 --> 00:37:13.320 Sadia Afrin Purba: we don't get any correlation here, because these two particulars, 0 and 1, these are correlated to each other. So, if the first qubit is 0, the second has to be 1. If the second is 1, the… 300 00:37:14.020 --> 00:37:17.439 Sadia Afrin Purba: This, this has to be zero, so this kind of correlation. 301 00:37:17.620 --> 00:37:19.830 Sadia Afrin Purba: Now, just moving to the paper. 302 00:37:20.630 --> 00:37:26.860 Sadia Afrin Purba: Okay, now, the main problem is the classical model faces a communication bottleneck. 303 00:37:26.980 --> 00:37:38.739 Sadia Afrin Purba: Where the internal height, latency dimension, or the hidden space, scale linearly with the number of sequence. For example, if we have… we want to, model a 304 00:37:39.020 --> 00:37:44.090 Sadia Afrin Purba: translation task. We have to, we have to have 305 00:37:44.260 --> 00:37:51.739 Sadia Afrin Purba: more hidden layer to capture it properly. So it's linear… linear with the number of samples. 306 00:37:52.080 --> 00:38:01.399 Sadia Afrin Purba: And how did they solve it? They used… they actually, came up with a very specific sequence translation task. 307 00:38:01.810 --> 00:38:10.400 Sadia Afrin Purba: And then they use entanglement to, solve the… Shoulder communication bottleneck. 308 00:38:11.200 --> 00:38:13.180 Sadia Afrin Purba: How did the endanglement shelf? 309 00:38:13.700 --> 00:38:28.380 Sadia Afrin Purba: And that non-correlation, non-local correlation provided by entangle substitute the need for explicit communication required to coordinate remote part of the input sequence. That is what I… giving… I was… 310 00:38:29.060 --> 00:38:38.660 Sadia Afrin Purba: I was giving a example that suppose we have a one entangled pair. I think we have a one entangled image. We know about it. 311 00:38:38.920 --> 00:38:41.660 Sadia Afrin Purba: So, we don't need to measure the other one. 312 00:38:41.990 --> 00:38:50.169 Sadia Afrin Purba: So that we don't need to do the, back and forth, or the information transfer… transformation. That is… 313 00:38:50.460 --> 00:38:53.650 Sadia Afrin Purba: How the entanglement helps that… helps us. 314 00:38:54.460 --> 00:38:59.699 Sadia Afrin Purba: Now, I really have to just quickly go through this game, the Magic Square game. 315 00:39:00.510 --> 00:39:02.510 Sadia Afrin Purba: Suppose we have two players. 316 00:39:02.730 --> 00:39:05.039 Sadia Afrin Purba: And there is no communication between them. 317 00:39:05.250 --> 00:39:17.629 Sadia Afrin Purba: And there is one referee. The referee will provide the two player, either 1, 2, 3, and the player will provide, some bit string of 3LIN. 318 00:39:18.170 --> 00:39:20.189 Sadia Afrin Purba: And the win condition is that 319 00:39:20.860 --> 00:39:29.500 Sadia Afrin Purba: Each row, the LS will provide the row string, so the each row must have auth parity, and even… and the column must be 320 00:39:29.620 --> 00:39:41.949 Sadia Afrin Purba: Event parity, sorry, row would be… row must be what parity, column must be event parity. And the intersecting cell has to… in the intersecting cell, the bit value should be same. 321 00:39:42.600 --> 00:39:52.569 Sadia Afrin Purba: So, it is… for example, the referee gave, X equal to 2 to the L is, so the row is 2, and Y is 1. Column is 1. And… 322 00:39:52.780 --> 00:39:56.969 Sadia Afrin Purba: Alice has to provide a, auth parity. 323 00:39:57.390 --> 00:40:06.560 Sadia Afrin Purba: So, he provides 010, and Bob provides 101. And in the intersecting, cell. 324 00:40:06.890 --> 00:40:09.980 Sadia Afrin Purba: The big string match, so they win. 325 00:40:10.470 --> 00:40:15.499 Sadia Afrin Purba: But, in this case, when the row and column is 2-2, 326 00:40:16.040 --> 00:40:19.310 Sadia Afrin Purba: The big string doesn't match, so they lose. 327 00:40:20.290 --> 00:40:24.459 Sadia Afrin Purba: So, if we play this game, mini-round, and fill up the… 328 00:40:24.570 --> 00:40:30.790 Sadia Afrin Purba: fill up all the shell in the classical way. Whatever base classical strategy you come up with. 329 00:40:31.000 --> 00:40:35.439 Sadia Afrin Purba: There is no chance you can do better than 8x9. 330 00:40:35.810 --> 00:40:47.950 Sadia Afrin Purba: In one case, you will always lose, but if you… if you use any quantum, state, or quantum strategy, then there is guarantee that you will win 100% time. 331 00:40:48.230 --> 00:40:54.130 Sadia Afrin Purba: So here, quantum strategy gives you, accuracy advantage. 332 00:40:54.890 --> 00:40:55.890 Sadia Afrin Purba: Negotion? 333 00:40:56.480 --> 00:41:01.450 Sadia Afrin Purba: And that's just because Alice and Bob communicate with each other using quantum entanglement. 334 00:41:01.580 --> 00:41:13.440 Sadia Afrin Purba: In quantum, in quantum case, they cannot even, they cannot communicate during the game, but before that, they can, come up, come up with a strategy. 335 00:41:14.010 --> 00:41:21.269 Sadia Afrin Purba: But during the game, they cannot, communicate. Okay, now, suppose… 336 00:41:21.530 --> 00:41:23.340 Sadia Afrin Purba: This one, yeah, it's a little mask. 337 00:41:23.470 --> 00:41:24.320 Sadia Afrin Purba: Okay. 338 00:41:24.590 --> 00:41:27.650 Sadia Afrin Purba: So, suppose, the… they… 339 00:41:27.760 --> 00:41:35.679 Sadia Afrin Purba: they come up with a quantum strategy, and they prepare a quantum state. This is one… one of the states. 340 00:41:36.070 --> 00:41:38.999 Sadia Afrin Purba: Okay, as the… one of the states. 341 00:41:39.990 --> 00:41:55.750 Sadia Afrin Purba: They plan that, yeah, we will measure our, measure our quantum state based on the row column we get. That is their strategy. They decided before the game. Now. 342 00:41:56.690 --> 00:42:02.309 Sadia Afrin Purba: For example, the referee gives, row 2 and Bob, column 3. 343 00:42:02.720 --> 00:42:04.980 Sadia Afrin Purba: Then… by measuring 344 00:42:05.340 --> 00:42:12.320 Sadia Afrin Purba: It is like we are rotating the quantum state in a X2 and Z3. This is just… 345 00:42:12.430 --> 00:42:18.240 Sadia Afrin Purba: some predefined basis, X3 and, J3. 346 00:42:18.950 --> 00:42:30.879 Sadia Afrin Purba: And we… and when we measure it, it collapsed to one state, one of the states, okay? We have this, superposition, state, and after measurement, we have this state. 347 00:42:31.230 --> 00:42:38.279 Sadia Afrin Purba: Now, the first two bits are for, Alice, and the last two bits are for Bob. 348 00:42:38.640 --> 00:42:51.240 Sadia Afrin Purba: Now, she saw that, she got 01, and the rule is she has to provide odd parity. So the third… so she knows that the third parity… third bit would be 0. 349 00:42:51.750 --> 00:43:00.910 Sadia Afrin Purba: Similarly, the Bob knows that, yeah, the first, first, two must be 1, 0, and the last has to be 1. 350 00:43:02.440 --> 00:43:07.380 Sadia Afrin Purba: And if they do that, if they follow their measurement outcome. 351 00:43:07.710 --> 00:43:14.459 Sadia Afrin Purba: they will always win. It is not only 2 and 3 case. For every row and column case, they will win. 352 00:43:14.630 --> 00:43:19.300 Sadia Afrin Purba: It is proven. So they will get 100% winning probability. 353 00:43:19.440 --> 00:43:23.869 Sadia Afrin Purba: If they use this kind of superposition and, entanglement stuff. 354 00:43:24.690 --> 00:43:26.909 Sadia Afrin Purba: Okay. But this cannot be done. 355 00:43:27.750 --> 00:43:31.750 Sadia Afrin Purba: No, not classical… with classical probability, you'll get… 356 00:43:33.060 --> 00:43:40.739 Sadia Afrin Purba: What makes it not possible to ask a problem? Because in classical wall, how can you… There is, 357 00:43:41.860 --> 00:43:56.719 Sadia Afrin Purba: I mean, there is no way that all the columns are odd probability, or odd parity, and all the rows are, even parity. I mean, in one of the cases, they lose. 358 00:43:57.500 --> 00:43:59.010 Sadia Afrin Purba: In the classical world. 359 00:43:59.570 --> 00:44:03.269 Sadia Afrin Purba: Or if you just design it using classical problems. Yeah. 360 00:44:03.430 --> 00:44:04.190 Sadia Afrin Purba: Yeah. 361 00:44:05.480 --> 00:44:06.160 Sadia Afrin Purba: Okay. 362 00:44:08.810 --> 00:44:15.739 Sadia Afrin Purba: So, if the game played, for example, 5 times, And Goldman Emmys… 363 00:44:15.920 --> 00:44:25.069 Sadia Afrin Purba: they have a state of one estate, and one quantum mistake other. So, as I know that one quantum state will collapse, as you do another. 364 00:44:25.240 --> 00:44:29.629 Sadia Afrin Purba: So, they will not get the information after one tax. 365 00:44:30.000 --> 00:44:32.389 Sadia Afrin Purba: How come they will continue this play? 366 00:44:34.390 --> 00:44:35.420 Sadia Afrin Purba: Oh. 367 00:44:37.790 --> 00:44:40.349 Sadia Afrin Purba: For every round, they will collapse them. 368 00:44:40.460 --> 00:44:47.650 Sadia Afrin Purba: quantum state, and get the beat value. Yeah. So they will need, for 5 times, they will need to organize it? 369 00:44:50.930 --> 00:45:02.900 Sadia Afrin Purba: But, the, we know the initial state. This is the initial state. So, we need one, one initial state. Five times, they probably used up, main, the same initial state. 370 00:45:05.560 --> 00:45:12.319 Sadia Afrin Purba: Okay, yeah, I got your point, that, for example. 371 00:45:12.500 --> 00:45:24.029 Sadia Afrin Purba: the… every time it will provide, it will get this way, but the randomness comes from the row and column value. Referee will provide you the, column and, 372 00:45:24.210 --> 00:45:25.630 Sadia Afrin Purba: Row number randomly. 373 00:45:26.220 --> 00:45:29.499 Sadia Afrin Purba: So, when you measure, the value would be different, everything. 374 00:45:31.120 --> 00:45:35.240 Sadia Afrin Purba: Because, the, measurement… Ms. Armin… 375 00:45:35.670 --> 00:45:41.679 Sadia Afrin Purba: operator, X2. X2, if you measure with X2, it will be different. 376 00:45:41.790 --> 00:45:49.089 Sadia Afrin Purba: If you measure with X3, to the same quantum, quantum state. It would be different, right? 377 00:45:49.280 --> 00:45:54.869 Sadia Afrin Purba: So the randomness comes from the row and column value that is given by the referee. 378 00:45:57.480 --> 00:46:00.320 Sadia Afrin Purba: Yeah, yeah, I know. 379 00:46:01.030 --> 00:46:03.210 Sadia Afrin Purba: I know, I know. Okay. 380 00:46:03.420 --> 00:46:19.109 Sadia Afrin Purba: Now, what they did, they take the concept of this magic rule and build a very specific sequence translation task. I really want to skip this part, but we can come back here. 381 00:46:20.810 --> 00:46:27.150 Sadia Afrin Purba: And they come up with a very specific role, not… not any of our traditional machine learning role. 382 00:46:27.260 --> 00:46:37.309 Sadia Afrin Purba: But it is also a, lining test. They are, lining some kind of sequence. Inputs… they are providing input and, predicting the output. 383 00:46:38.730 --> 00:46:39.610 Sadia Afrin Purba: Okay. 384 00:46:43.230 --> 00:46:47.640 Sadia Afrin Purba: Now, now the main outcome is… 385 00:46:47.800 --> 00:46:55.589 Sadia Afrin Purba: But by doing that, the main outcome is, if the same task they perform, classically, the… 386 00:46:56.120 --> 00:47:01.570 Sadia Afrin Purba: accuracy, or the score will never be 100%. But with quantum strategy. 387 00:47:01.680 --> 00:47:05.380 Sadia Afrin Purba: They get 100% accuracy, 100% score. 388 00:47:06.950 --> 00:47:11.200 Sadia Afrin Purba: So, and based on their findings, they provide, three 389 00:47:11.350 --> 00:47:14.550 Sadia Afrin Purba: theorem. The first theorem established that, 390 00:47:14.890 --> 00:47:19.330 Sadia Afrin Purba: During inference, we get exponential separation. 391 00:47:19.540 --> 00:47:23.929 Sadia Afrin Purba: in complexity. Because, in inference. 392 00:47:23.990 --> 00:47:32.190 Sadia Afrin Purba: We have to suppose the next token prediction. To predict the next token, we have to iterate all our previous token, right? 393 00:47:32.210 --> 00:47:44.009 Sadia Afrin Purba: In classical world, so it is linear. But in quantum world, it has, it, it will not grow, within a number of, number of, number of sequence. It will, it is constant. 394 00:47:45.130 --> 00:47:49.139 Sadia Afrin Purba: And another one is, they, they… 395 00:47:49.360 --> 00:48:01.970 Sadia Afrin Purba: provide the inference advantage with constant noise. So, they use constant noise and see that, yeah, they are advantageous. The O , the constant, advantage still remains. 396 00:48:02.710 --> 00:48:03.530 Sadia Afrin Purba: And… 397 00:48:03.640 --> 00:48:17.230 Sadia Afrin Purba: Same training identities. For training, algorithm, we need… if we have n-dimensional data, we need n-dimensional hidden layer to capture it… capture the perfect, output. 398 00:48:18.050 --> 00:48:26.059 Sadia Afrin Purba: But, for quantum, we need a constant, requirement. This is a quantum side, circuit. 399 00:48:26.440 --> 00:48:35.030 Sadia Afrin Purba: Now, here is a, little numerical, demonstration that if we used a quantum 400 00:48:35.030 --> 00:48:38.870 Sadia Afrin Purba: Quantum algorithm, we would need only 1.28 parameter. 401 00:48:38.870 --> 00:48:55.650 Sadia Afrin Purba: But if we use any classical algorithm, any, recurrent algorithm or encoder-decoder, there is, like, 3, 3 into 10 to the power 4, or 6 into 10 to the power 4 parameter. And by using quantum algorithm, we are actually, getting 100% 402 00:48:55.950 --> 00:48:59.630 Sadia Afrin Purba: Accuracy. So, parameter-wise, accuracy-wise, it is good. 403 00:49:00.320 --> 00:49:08.340 Sadia Afrin Purba: These are their experimentation results, so the main outcome is the quantum model achieved a perfect score for this particular task. 404 00:49:09.080 --> 00:49:16.759 Sadia Afrin Purba: about 2 to 3 order of magnitude for your trainable parameter than the classical algorithm. That is the main 405 00:49:17.080 --> 00:49:18.770 Sadia Afrin Purba: Dish of this paper. 406 00:49:20.020 --> 00:49:20.750 Sadia Afrin Purba: Okay. 407 00:49:21.010 --> 00:49:25.450 Sadia Afrin Purba: Do you have any 408 00:49:26.300 --> 00:49:28.340 Sadia Afrin Purba: But we only got the 5 minutes. 409 00:49:28.660 --> 00:49:33.100 Sadia Afrin Purba: The webinar queue tried to summarize the other two papers quickly. 410 00:49:33.230 --> 00:49:37.120 Sadia Afrin Purba: Okay. And then we'll continue some of this… The private sector. 411 00:49:38.600 --> 00:49:39.330 Sadia Afrin Purba: Thanks. 412 00:49:40.650 --> 00:49:46.780 Sadia Afrin Purba: Okay, the second… the entangled data and quantum intervals. The main idea is 413 00:49:47.120 --> 00:49:59.880 Sadia Afrin Purba: We need only… if we have an entangled dataset, if we have that, and we want to learn anything from that entangled dataset, we only need one 414 00:50:00.010 --> 00:50:03.000 Sadia Afrin Purba: Training, training point. 415 00:50:03.740 --> 00:50:05.770 Sadia Afrin Purba: To learn the system. 416 00:50:06.030 --> 00:50:07.519 Sadia Afrin Purba: That is the main summary. 417 00:50:09.280 --> 00:50:16.060 Sadia Afrin Purba: Okay, and they provide a theory. This is the generational risk theory. 418 00:50:20.820 --> 00:50:24.399 Sadia Afrin Purba: And it is also a numerical thing, for example. 419 00:50:24.590 --> 00:50:28.690 Sadia Afrin Purba: We have one, one training data, and… 420 00:50:28.920 --> 00:50:35.310 Sadia Afrin Purba: the system is maximally entangled. Then the law, then the risk is minus. 421 00:50:35.940 --> 00:50:42.489 Sadia Afrin Purba: But the quantum risk is non-negative by definition. So, when we get a minus value. 422 00:50:42.650 --> 00:50:45.340 Sadia Afrin Purba: It is… means we are getting a zero value. 423 00:50:45.490 --> 00:50:46.190 Sadia Afrin Purba: Okay. 424 00:50:47.060 --> 00:50:49.119 Sadia Afrin Purba: This is a circuit they use. 425 00:50:49.900 --> 00:50:53.999 Sadia Afrin Purba: And, they, simulate their result. 426 00:50:56.120 --> 00:51:04.789 Sadia Afrin Purba: Now, there are many open questions. For example, this particular system actually use, assume that we have 427 00:51:04.970 --> 00:51:09.359 Sadia Afrin Purba: a perfect training condition. It does not happen like that way, right? 428 00:51:09.910 --> 00:51:14.420 Sadia Afrin Purba: And the third paper, In the third paper, 429 00:51:14.790 --> 00:51:28.839 Sadia Afrin Purba: They use quantum disc code, not the quantum, quantum entanglement, because quantum discord can provide some, correlation, and they use that correlation to get, advantage. 430 00:51:31.760 --> 00:51:35.819 Sadia Afrin Purba: Okay, they provide, they say that if there is no covariance. 431 00:51:35.990 --> 00:51:39.679 Sadia Afrin Purba: Then, the, kernel system, the… 432 00:51:40.300 --> 00:51:54.949 Sadia Afrin Purba: carnal machines cannot learn anything. That is their one theory. So, if we have a coherent quantum system, the learning performance of a kernel function, we can learn a kernel function. 433 00:51:55.700 --> 00:52:04.830 Sadia Afrin Purba: And another theory is, if we have a coherent quantum system, we can generate some, this… that is another theory. 434 00:52:05.990 --> 00:52:09.140 Sadia Afrin Purba: So, this is the technical thing. 435 00:52:10.420 --> 00:52:13.170 Sadia Afrin Purba: And this is, they actually show that 436 00:52:13.480 --> 00:52:18.960 Sadia Afrin Purba: Alpha is a purity value. If they improve the purity, the accuracy also improves. 437 00:52:21.820 --> 00:52:26.460 Sadia Afrin Purba: But there are many open questions. For example, there are lack of 438 00:52:26.520 --> 00:52:42.770 Sadia Afrin Purba: comparison between, between the classical algorithm and their algorithm. Also, there are many, quantum algorithms that, classify… that is used to classify the carnal metrics. They didn't compare those algorithms together. 439 00:52:43.060 --> 00:52:46.359 Sadia Afrin Purba: So, there are many open questions about that. 440 00:52:46.820 --> 00:52:57.220 Sadia Afrin Purba: So, in summary, we have this, that if we use entanglement, there is possible that we will get computational speedup and data efficiency. 441 00:52:57.660 --> 00:53:03.629 Sadia Afrin Purba: And… The second one is theoretical and sensitive to noise, but… 442 00:53:04.470 --> 00:53:07.619 Sadia Afrin Purba: It is a very important theorem. 443 00:53:07.850 --> 00:53:10.370 Sadia Afrin Purba: And if we use, this code. 444 00:53:10.490 --> 00:53:15.159 Sadia Afrin Purba: Then we can get some advantages at present time. 445 00:53:16.030 --> 00:53:31.620 Sadia Afrin Purba: So, but the most of the proof assume the script data. For example, in the first paper, they used 01-bit string only, and assume perfect training condition, and constant device noise, but in practice, we deal with, 446 00:53:31.790 --> 00:53:40.300 Sadia Afrin Purba: continuous data training error. So, there is a huge gap in the in their, theorem. 447 00:53:40.470 --> 00:53:47.279 Sadia Afrin Purba: So we need a quantum research roadmap that is theoretically complete, measurable, and deployable. 448 00:53:47.410 --> 00:53:51.590 Sadia Afrin Purba: So, for that, I have some roadmap. For example, we need 449 00:53:51.860 --> 00:53:58.910 Sadia Afrin Purba: At least two theories that actually reformed the quantum freelance theory. 450 00:53:59.470 --> 00:54:02.300 Sadia Afrin Purba: I mean, we have… we need a learning theorem. 451 00:54:02.480 --> 00:54:06.690 Sadia Afrin Purba: That also… that also consider imperfect training. 452 00:54:06.820 --> 00:54:12.639 Sadia Afrin Purba: Condition, and also another, theorem that, consider 453 00:54:13.260 --> 00:54:20.620 Sadia Afrin Purba: All, all machine learning tasks, not only, not only a specific handmade machine learning task. 454 00:54:21.330 --> 00:54:29.290 Sadia Afrin Purba: And we need a proper data encoding, and try to figure out how data encoding, in, data encoding. 455 00:54:30.020 --> 00:54:35.439 Sadia Afrin Purba: Effect the quantum advantage, and evaluation metric. 456 00:54:35.680 --> 00:54:47.869 Sadia Afrin Purba: For classical one, we need, we use p-value or statistical significance. But in quantum world, if we only rely on the p-value, we actually are not going to… 457 00:54:48.050 --> 00:55:05.660 Sadia Afrin Purba: going to get the, proper evaluation. We need to consider the, wall clock value. We also need to consider that quantum device might use less amount of energy, rather than the high-performance computing. 458 00:55:05.800 --> 00:55:12.159 Sadia Afrin Purba: And… We have to consider those, not only the p-value and all the… and all the accuracy thing. 459 00:55:12.750 --> 00:55:18.190 Sadia Afrin Purba: And then we don't have any proper benchmark data set. First of all, we need entangle dataset. 460 00:55:18.770 --> 00:55:23.009 Sadia Afrin Purba: We… we can generate integral datasheet from the lab, but 461 00:55:23.140 --> 00:55:37.220 Sadia Afrin Purba: there is a… there should be a good recess field where we try to find out where is the entangled data, how can we find the entangled data, or how can we convert the classical data to entanglement? And… 462 00:55:37.490 --> 00:55:40.930 Sadia Afrin Purba: As a classical dataset, most of the literature use MNIST 463 00:55:41.060 --> 00:55:44.749 Sadia Afrin Purba: classical data. This is very easy data for 464 00:55:44.920 --> 00:55:48.570 Sadia Afrin Purba: Classical problem. We need some more difficult problem. 465 00:55:48.890 --> 00:55:52.630 Sadia Afrin Purba: To have a proper benchmark result. 466 00:55:52.950 --> 00:56:05.220 Sadia Afrin Purba: And then, we have to build a system, a hybrid system, where a quantum and classical will be, tied together. And here, also, we can, 467 00:56:05.380 --> 00:56:19.350 Sadia Afrin Purba: we can mix up the different kind of correlation. For example, we can use the entanglement for the expressive model, and this code for a noise-resilient system. 468 00:56:19.840 --> 00:56:30.200 Sadia Afrin Purba: And then the pilot domain could be an easy signal segmentation, because it is a very difficult task in a, classical setting. 469 00:56:30.200 --> 00:56:45.299 Sadia Afrin Purba: Or multimodal classification. Multimodal classification, as I mentioned, because if we want to, classify all the breed of the dog, we need trillions of data. We can minimize that if we use the quantum entanglement and quantum property. 470 00:56:45.550 --> 00:56:47.650 Sadia Afrin Purba: So, these are the roadmap. 471 00:56:48.570 --> 00:56:52.730 Sadia Afrin Purba: And I want to close my thought that 472 00:56:53.840 --> 00:56:57.120 Sadia Afrin Purba: We… why do we want to do machine learning? 473 00:56:57.240 --> 00:57:04.779 Sadia Afrin Purba: Many… because we want to solve the intelligence, we want to understand our… how our brain works, and all that stuff. 474 00:57:05.070 --> 00:57:22.780 Sadia Afrin Purba: But, only focusing on that, we forget about that there are other intelligents, other intelligents from the animal. In fact, there are some intelligent in the plant. For example, the leaves move to a certain direction to get the sun and all that stuff. 475 00:57:23.200 --> 00:57:27.900 Sadia Afrin Purba: And if I look into this video, I, 476 00:57:28.180 --> 00:57:34.780 Sadia Afrin Purba: It is using bread as a trap, and didn't catch the big fish, because it cannot carry it. 477 00:57:35.030 --> 00:57:39.350 Sadia Afrin Purba: But, it will carry, it will catch the small fish. 478 00:57:39.860 --> 00:57:48.880 Sadia Afrin Purba: So, I think it is a very intelligent BART, and if we provide… if we create a robot bart with lots of data, still. 479 00:57:49.070 --> 00:57:53.699 Sadia Afrin Purba: That part can… that robot bird cannot give such performance. 480 00:57:54.190 --> 00:58:06.769 Sadia Afrin Purba: And the quantum system, the quantum particle is universal. So, to get a artificial general intellect, we really should focus on this quantum 481 00:58:07.280 --> 00:58:09.520 Sadia Afrin Purba: Machine learning mechanism. 482 00:58:09.860 --> 00:58:28.720 Sadia Afrin Purba: And also, recently, we have figured out that in our brain neuron, we, we can find out quantum properties. So it is a very miracle that in the noisy environment, there are… there can be quantum effect. So… 483 00:58:29.310 --> 00:58:38.700 Sadia Afrin Purba: This is very… This has many more potential, I think. And this is the references. 484 00:58:39.390 --> 00:58:40.360 Sadia Afrin Purba: Thank you. 485 00:58:45.760 --> 00:58:50.199 Sadia Afrin Purba: In the interest of time, I will excuse the… Now, committee members? 486 00:58:51.640 --> 00:58:54.110 Sadia Afrin Purba: Continuous. 487 00:58:57.320 --> 00:58:58.139 Sadia Afrin Purba: Is it coming? 488 00:59:09.200 --> 00:59:10.170 Sadia Afrin Purba: Hammers? 489 00:59:12.390 --> 00:59:21.439 Sadia Afrin Purba: No, actually. I… Yeah, fucking stop, work.