Scribd
Upload
10 views

Module7 Qgans

This document discusses using quantum machine learning techniques for generative modeling tasks. Specifically, it proposes using a quantum generative adversarial network (QGAN) for approximate quantum data loading. A QGAN would involve a quantum generator circuit creating potential data instances, and a discriminator circuit determining if instances are similar to real training data or generated samples. This approach could provide enhanced expressiveness over classical models and enable exploitation of quantum advantages for generative machine learning applications.

Uploaded by

dr.hassanalqanazi
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
10 views

Module7 Qgans

This document discusses using quantum machine learning techniques for generative modeling tasks. Specifically, it proposes using a quantum generative adversarial network (QGAN) for approximate quantum data loading. A QGAN would involve a quantum generator circuit creating potential data instances, and a discriminator circuit determining if instances are similar to real training data or generated samples. This approach could provide enhanced expressiveness over classical models and enable exploitation of quantum advantages for generative machine learning applications.

Uploaded by

dr.hassanalqanazi
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 13

Quantum Machine Learning

Approximate Data Loading with Quantum


Generative Adversarial Networks

Julien Gacon & Dr. Christa Zoufal


Quantum Applications Researchers
Generative Machine Learning Change picture

Idea
Characterize data generation of given data samples

Approach
Train parameterized models to approximate the underlying probability
distribution

Applications
Data generation, data analysis, etc.
Generative Quantum Machine Learning
Idea
Equip generative models with quantum resources that are potentially hard
to simulate classically

Improvement
Known: enhanced expressive power [1, 2]
Expected: trains faster or with less resources [2]

Quantum Application
Efficient, approximate quantum data loading
è To enable the exploitation of potential quantum advantage

[1] Y. Du, M.-H. Hsieh, T. Liu, and D. Tao. Expressive power of parametrized quantum circuits.
Physical Review Research, 2(3), 2020
[2] S. Lloyd and C. Weedbrook. Quantum generative adversarial learning. Phys. Rev.Lett., 121, 2018
Approximate Quantum Data Loading

p(x)

Generally exponentially expensive

x
Approximate Quantum Data Loading

p(x)

x
Approximate Quantum Data Loading

p(x)

Amplitude Encoding
P1 p
2n
<latexit sha1_base64="D/PqTLeVBmK85Nh6e866i/0uc/U=">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</latexit>

px (!) |xi
b=0

x
Approximate Quantum Data Loading

p(x)

Avoid exponential overhead with


approximate loading

x
(a) (b)

Quantum Generative Adversarial Networks

(a) (c) (d)

(e) (f)
Idea: Train parameters of a variational quantum (b)
circuit for quantum data loading
Architecture
training data samples
generated data samples x!l!
<latexit sha1_base64="QbHbB0H+7Ul59sRfBXbUf8wKpls=">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</latexit>

g"l #
"<latexit sha1_base64="XWEFS9hrsEkSUqcbxUXfwhMhHgA=">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</latexit>

Generator
!!
Discriminator classification:
"" real / fake

Generator: Create data instances that have similar properties as the training data
Discriminator: Distinguish between training data and generated data
Quantum Generator

Sampling probability of |bi

measure map
|ji gj

Statistical information about the training data can guide the choice for
Training

P
m
<latexit sha1_base64="fhPuzYl0wk23Gl3Hj7A7DsNZuFw=">AAAC73icbVFdaxNBFJ1dv9pUbaqPvgymhRQ0ZIOoKIWiLfqgWMG0gUy6zE5uNtPM7Kwzs8WwzO/wTXzxwZ/kv3E2WcQ0vTDcw7kf5947SS64sd3unyC8cfPW7Tsbm42tu/fubzd3HpwaVWgGfaaE0oOEGhA8g77lVsAg10BlIuAsmb2t4meXoA1X2Rc7z2EkaZrxCWfUeipu/trdxB/id0TAxLZJosTYzKV3Jcmn3D3BK5SSkFJHNE+ndh8f4KfEFJIILrk1cXlxELnzUro8vmhfU7ZPhEqXOvjofNl/XTZ1vtE/idrvxs1Wt9NdGF4HUQ1aqLaTeCfokbFihYTMMkGNGUbd3I5Kqi1nAlyDFAZyymY0haGHGZVgRuXing7veWaMJ0r7l1m8YP+vKKk01bw+U1I7NVdjFXldbFjYyctRybO8sJCxpdCkENgqXH0OHnMNzIq5B5Rp7mfFbEo1ZdZ/4YpKoukMrFtVZsyvazx5BH5pDR/9AJ9y0NQqXZLjgStJNVOSlMfO7WH4lns1v+klFQU0Gv7K0dWbroPTXid63nn2udc6fFPfewM9Qo9RG0XoBTpE79EJ6iMWbAVR8Cp4HX4Nv4c/wp/L1DCoax6iFQt//wWk9+8w</latexit>

LG ( , !) = pj (!) log D gj
j=1

Sampling
probability of
<latexit sha1_base64="DqR8qNgRKhjYlfhNBY4eWQEN168=">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</latexit>

gj

P
m
<latexit sha1_base64="Nh5jvmOinYTSCZe6Jp0OmIE+fhA=">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</latexit>

1
LD ( , !) = m log D (xj ) + pj (!) log 1 D gj
j=1

Number of data Training data


samples in batch samples

Maximize 𝐿! (𝛟, 𝝎) w.r.t. 𝝎 to generate data samples that look like training data samples
Maximize 𝐿" (𝛟, 𝝎) w.r.t. 𝛟 to discriminate between training data samples and generated data samples
Illustrative Examples

(e) (f)

Simulation Hardware
To sum up

• Demonstrate variational implementation on quantum hardware

• Present a thorough benchmarking study including multi-modal and multivariate

distributions

• Illustrate approximate quantum data loading for quantum-enhanced option pricing

You might also like