microsoft research academic services gtm draftย ยท training quantum boltzmann machines we train...
TRANSCRIPT
Nathan WiebeWork with Maria Kieferova
Quantum Neural Networks
Machine learning
โข Fast scalable quantum computers will fundamentally change machine learning.
โข Quadratic (or greater) speedups.
โข Better models for data.
โข Improved privacy and security.
Quantum Machine Learning
โฎ
โข Fast scalable quantum computers will fundamentally change machine learning.
โข Quadratic (or greater) speedups.
โข Better models for data.
โข Improved privacy and security.
Quantum Machine Learning
Quantum Boltzmann Machine Training.
โฎ
Boltzmann Machines
Class of recurrent neural network that is strongly related to physics.
๐ ๐ฅ โ ๐โ๐ธ(๐ฅ)
1 = cat
0 = Not cat
Energy
, 1
, 0
, 1
, 0
๐ธ ๐ฅ =
<๐,๐>
๐ค๐,๐๐๐๐๐
๐ธ ๐ฅ โ ๐ป =
<๐,๐>
๐ค๐,๐ 1 โ ๐๐ 1 โ ๐๐4
๐ค = ๐๐๐๐๐๐๐ค(๐ท๐พ๐ฟ(๐ ๐ฅ |๐๐ก๐๐๐๐(๐ฅ)))
Quantizing the Boltzmann Machine
๐ โ ๐โ๐ป
๐ป =
๐
๐ค๐๐ป๐
๐ค = ๐๐๐๐๐๐๐ค ๐(๐|๐๐ก๐๐๐๐)
Quantum Boltzmann machineslearn directly from quantum states.
They also generate quantum statesthat closely resemble the training data.
๐ป๐Do not commute
Energy
, 1
, 0
, 1
, 0
Classically we have only energy penalties.
Quantumly, we controlboth the energiesand the eigenbasis theenergies are applied in.
Classical Quantum
Training quantum Boltzmann Machines
We train using gradient descent on the quantum relative entropy (quantum KL divergence).
๐ =๐โ ฯ๐๐ค๐๐ป๐
๐๐(๐โฯ๐๐ค๐๐ป๐)
log ๐ = โฯ๐๐ค๐๐ป๐ โ log ๐๐(๐โ ฯ๐๐ค๐๐ป๐
๐๐ค๐log ๐ = โ๐ป๐ + ๐๐ ๐๐ป๐
๐๐ค๐๐ ๐ ๐ = โTr ๐๐ป๐ + ๐๐ ๐๐ป๐ = ๐ป๐ ๐๐๐๐๐
โ ๐ป๐ ๐๐๐ก๐
The number of times the trainingdata is accessed and Gibbs states
are prepared scales as ๐๐2
๐2.
๐ = number of weights.
Relative entropy training
Train
ing
ob
jectiv
e fu
nctio
n
Learning a model for a quantum system
Assume a quantum Hamiltonian thatis a transverse Ising model:
Find, ๐ผ, ๐ฝ and ๐พ thatbest represent the Hamiltonian giventraining data
๐ = ๐โเทฉ๐ป/๐
Want to minimize ฮ ๐ป = || ๐ป โ เทฉ๐ป||
Building a quantum associative memory
Quantum Training vectors: ๐ฃ๐ = ๐๐ โ๐๐ Data qubits
0
0
Address qubits
Pick ๐ป๐ from a universal set of Hamiltonians on 6 qubits. 1000 epochs, learning rate=0.2
Relative entropy โ 10โ3.
Conclusions
Quantum computers enable new forms of machine learning.
Quantum Neural networks can learn directly from quantum data.
This opens up much richer models of data that cannot be efficiently trained using classical computers.
What is machine learning?Machine learning uses computers to find patterns in data,classify data, or perform tasks.
Computers are not usually told explicitly how to do this, or what โfeaturesโ to use.
4 0 5
โฏ
Training Data
???
4
ยฉ 2014 Microsoft Corporation. All rights reserved.