quantum mechanics, tensor networks and machine learning
TRANSCRIPT
![Page 1: Quantum mechanics, tensor networks and machine learning](https://reader034.vdocuments.net/reader034/viewer/2022042311/625b6ac2a848f53f1a4e5b32/html5/thumbnails/1.jpg)
Quantum mechanics, tensor networks and machine learning
Nicola Pancotti, quantum research scientist @ Amazon Web Services
![Page 2: Quantum mechanics, tensor networks and machine learning](https://reader034.vdocuments.net/reader034/viewer/2022042311/625b6ac2a848f53f1a4e5b32/html5/thumbnails/2.jpg)
in collaboration with
● Max Planck Quantum Optics: Ivan Glasser, J. Ignacio Cirac, Ivan Diego Rodriguez
● Technical University Munich: Moritz August
● Free University Berlin:Ryan Sweke, Jens Eisert
![Page 3: Quantum mechanics, tensor networks and machine learning](https://reader034.vdocuments.net/reader034/viewer/2022042311/625b6ac2a848f53f1a4e5b32/html5/thumbnails/3.jpg)
motivations
![Page 4: Quantum mechanics, tensor networks and machine learning](https://reader034.vdocuments.net/reader034/viewer/2022042311/625b6ac2a848f53f1a4e5b32/html5/thumbnails/4.jpg)
● Quantum mechanics and machine learning are intrinsically probabilistic theories
● Neural networks and tensor networks are two extremely successful paradigms in their respective fields
● Can we connect their mathematical formulation?● Can we improve one by using the other?
![Page 5: Quantum mechanics, tensor networks and machine learning](https://reader034.vdocuments.net/reader034/viewer/2022042311/625b6ac2a848f53f1a4e5b32/html5/thumbnails/5.jpg)
Pattern Recognition
Machine Learning Tensor Networks
Classical Data Quantum Data
![Page 6: Quantum mechanics, tensor networks and machine learning](https://reader034.vdocuments.net/reader034/viewer/2022042311/625b6ac2a848f53f1a4e5b32/html5/thumbnails/6.jpg)
content of the talk
![Page 7: Quantum mechanics, tensor networks and machine learning](https://reader034.vdocuments.net/reader034/viewer/2022042311/625b6ac2a848f53f1a4e5b32/html5/thumbnails/7.jpg)
1. quantum mechanics and linear algebraa. wavefunctions are vectorsb. observables are matrices
2. linear algebra and tensor networks a. Tensor networks as an efficient tool for certain
problems in linear algebra3. tensor networks and machine learning
a. tensor networks for probabilistic modeling b. Examples for supervised and unsupervised learning
![Page 8: Quantum mechanics, tensor networks and machine learning](https://reader034.vdocuments.net/reader034/viewer/2022042311/625b6ac2a848f53f1a4e5b32/html5/thumbnails/8.jpg)
quantum mechanics
![Page 9: Quantum mechanics, tensor networks and machine learning](https://reader034.vdocuments.net/reader034/viewer/2022042311/625b6ac2a848f53f1a4e5b32/html5/thumbnails/9.jpg)
wavefunctions are vectors
A generic vector in this Hilbert space can be expressed as:
with
![Page 10: Quantum mechanics, tensor networks and machine learning](https://reader034.vdocuments.net/reader034/viewer/2022042311/625b6ac2a848f53f1a4e5b32/html5/thumbnails/10.jpg)
observables are matrices
Dynamical behavior: Schrödinger equation
energy
magnetizationtwo examples
![Page 11: Quantum mechanics, tensor networks and machine learning](https://reader034.vdocuments.net/reader034/viewer/2022042311/625b6ac2a848f53f1a4e5b32/html5/thumbnails/11.jpg)
tensor networks and linear algebra
![Page 12: Quantum mechanics, tensor networks and machine learning](https://reader034.vdocuments.net/reader034/viewer/2022042311/625b6ac2a848f53f1a4e5b32/html5/thumbnails/12.jpg)
matrix product representations
● physical degrees of freedom are arranged on a line: one dimension.
● the vector fulfills an area law
in physics, if
then matrix product states are a faithful representation
![Page 13: Quantum mechanics, tensor networks and machine learning](https://reader034.vdocuments.net/reader034/viewer/2022042311/625b6ac2a848f53f1a4e5b32/html5/thumbnails/13.jpg)
graphical notationRank 2 tensors at the edges Rank 3 tensors in the bulk
Efficient matrix-vector multiplication
𝜮
![Page 14: Quantum mechanics, tensor networks and machine learning](https://reader034.vdocuments.net/reader034/viewer/2022042311/625b6ac2a848f53f1a4e5b32/html5/thumbnails/14.jpg)
tensor networks and machine learning
![Page 15: Quantum mechanics, tensor networks and machine learning](https://reader034.vdocuments.net/reader034/viewer/2022042311/625b6ac2a848f53f1a4e5b32/html5/thumbnails/15.jpg)
Matrix Product States (MPS) :
String Bond States (SBS):
Boltzmann Machines (BM):
Restricted Boltzmann Machines (RBM):
PRX 8 (1), 011006
Restricted Boltzmann Machines are a subclass of string bond states
![Page 16: Quantum mechanics, tensor networks and machine learning](https://reader034.vdocuments.net/reader034/viewer/2022042311/625b6ac2a848f53f1a4e5b32/html5/thumbnails/16.jpg)
Relationship with graphical modelsGraphical models are classical probabilistic models where one assumes a certain factorization of the probability density function
Without hidden units:
With hidden units:
![Page 17: Quantum mechanics, tensor networks and machine learning](https://reader034.vdocuments.net/reader034/viewer/2022042311/625b6ac2a848f53f1a4e5b32/html5/thumbnails/17.jpg)
Locality of the RBMlocal connections with local connections
IEEE Access 8, 68169-68182
![Page 18: Quantum mechanics, tensor networks and machine learning](https://reader034.vdocuments.net/reader034/viewer/2022042311/625b6ac2a848f53f1a4e5b32/html5/thumbnails/18.jpg)
Combining different models
![Page 19: Quantum mechanics, tensor networks and machine learning](https://reader034.vdocuments.net/reader034/viewer/2022042311/625b6ac2a848f53f1a4e5b32/html5/thumbnails/19.jpg)
classification
![Page 20: Quantum mechanics, tensor networks and machine learning](https://reader034.vdocuments.net/reader034/viewer/2022042311/625b6ac2a848f53f1a4e5b32/html5/thumbnails/20.jpg)
Image Classification Goal: Given a dataset of images
and corresponding labels, we want to predict the label of a new image
![Page 21: Quantum mechanics, tensor networks and machine learning](https://reader034.vdocuments.net/reader034/viewer/2022042311/625b6ac2a848f53f1a4e5b32/html5/thumbnails/21.jpg)
Image Classification Goal: Given a dataset of images
and corresponding labels, we want to predict the label of a new image
Choose a ”model” :
Define a cost function :
![Page 22: Quantum mechanics, tensor networks and machine learning](https://reader034.vdocuments.net/reader034/viewer/2022042311/625b6ac2a848f53f1a4e5b32/html5/thumbnails/22.jpg)
FashionMNISTSVM 84.1%
Multilayer Perceptron
87.7%
SBS 89.0%
AlexNet 89.9%
1-layer CNN+SBS 92.3%
GoogLeNet 93.7%
IEEE Access 8, 68169-68182
![Page 23: Quantum mechanics, tensor networks and machine learning](https://reader034.vdocuments.net/reader034/viewer/2022042311/625b6ac2a848f53f1a4e5b32/html5/thumbnails/23.jpg)
Architecture
IEEE Access 8, 68169-68182
![Page 24: Quantum mechanics, tensor networks and machine learning](https://reader034.vdocuments.net/reader034/viewer/2022042311/625b6ac2a848f53f1a4e5b32/html5/thumbnails/24.jpg)
maximum likelihood estimations
![Page 25: Quantum mechanics, tensor networks and machine learning](https://reader034.vdocuments.net/reader034/viewer/2022042311/625b6ac2a848f53f1a4e5b32/html5/thumbnails/25.jpg)
Maximum likelihood
Learn from a database:
Learn from a distribution:
![Page 26: Quantum mechanics, tensor networks and machine learning](https://reader034.vdocuments.net/reader034/viewer/2022042311/625b6ac2a848f53f1a4e5b32/html5/thumbnails/26.jpg)
Some models for unsupervised learning
Tensor Train (MPS):
Born Machines:
Locally Purified States:
![Page 27: Quantum mechanics, tensor networks and machine learning](https://reader034.vdocuments.net/reader034/viewer/2022042311/625b6ac2a848f53f1a4e5b32/html5/thumbnails/27.jpg)
Expressive power
NeurIPS, 2019
![Page 28: Quantum mechanics, tensor networks and machine learning](https://reader034.vdocuments.net/reader034/viewer/2022042311/625b6ac2a848f53f1a4e5b32/html5/thumbnails/28.jpg)
practical applications: random distributions
NeurIPS, 2019
![Page 29: Quantum mechanics, tensor networks and machine learning](https://reader034.vdocuments.net/reader034/viewer/2022042311/625b6ac2a848f53f1a4e5b32/html5/thumbnails/29.jpg)
practical applications: real data sets
NeurIPS, 2019
![Page 30: Quantum mechanics, tensor networks and machine learning](https://reader034.vdocuments.net/reader034/viewer/2022042311/625b6ac2a848f53f1a4e5b32/html5/thumbnails/30.jpg)
conclusions
![Page 31: Quantum mechanics, tensor networks and machine learning](https://reader034.vdocuments.net/reader034/viewer/2022042311/625b6ac2a848f53f1a4e5b32/html5/thumbnails/31.jpg)
● quantum mechanics and linear algebra● linear algebra and tensor networks ● graphical models can be mapped to tensor
networks ● tensor networks can be used for
○ classification problems○ modeling probabilistic theories
● tensor networks can provide deeper mathematical insights
![Page 32: Quantum mechanics, tensor networks and machine learning](https://reader034.vdocuments.net/reader034/viewer/2022042311/625b6ac2a848f53f1a4e5b32/html5/thumbnails/32.jpg)
thanks for your attention