nathan killoran - quantum computer · 2019. 7. 31. · nathan killoran. quantum computers are good...
TRANSCRIPT
![Page 1: Nathan Killoran - Quantum Computer · 2019. 7. 31. · Nathan Killoran. Quantum computers are good at: Optimization Linear algebra Sampling Quantum physics. Quantum Machine Learning](https://reader033.vdocuments.net/reader033/viewer/2022051913/60043dc775bf5a42c606782e/html5/thumbnails/1.jpg)
Quantum Machine Learning
Nathan Killoran
![Page 2: Nathan Killoran - Quantum Computer · 2019. 7. 31. · Nathan Killoran. Quantum computers are good at: Optimization Linear algebra Sampling Quantum physics. Quantum Machine Learning](https://reader033.vdocuments.net/reader033/viewer/2022051913/60043dc775bf5a42c606782e/html5/thumbnails/2.jpg)
Quantum computers are good at:
OptimizationLinear algebra SamplingQuantum physics
![Page 3: Nathan Killoran - Quantum Computer · 2019. 7. 31. · Nathan Killoran. Quantum computers are good at: Optimization Linear algebra Sampling Quantum physics. Quantum Machine Learning](https://reader033.vdocuments.net/reader033/viewer/2022051913/60043dc775bf5a42c606782e/html5/thumbnails/3.jpg)
Quantum Machine Learning papers
![Page 4: Nathan Killoran - Quantum Computer · 2019. 7. 31. · Nathan Killoran. Quantum computers are good at: Optimization Linear algebra Sampling Quantum physics. Quantum Machine Learning](https://reader033.vdocuments.net/reader033/viewer/2022051913/60043dc775bf5a42c606782e/html5/thumbnails/4.jpg)
Quantum Machine Learning• AI/ML already uses special-purpose processors: GPUs, TPUs, ASICs
• Quantum computers (QPUs) could be used as special-purpose AI accelerators
• May enable training of previously intractable models
![Page 5: Nathan Killoran - Quantum Computer · 2019. 7. 31. · Nathan Killoran. Quantum computers are good at: Optimization Linear algebra Sampling Quantum physics. Quantum Machine Learning](https://reader033.vdocuments.net/reader033/viewer/2022051913/60043dc775bf5a42c606782e/html5/thumbnails/5.jpg)
New AI models• Quantum computing can also lead to
new machine learning models
• Examples currently being studied are:
- Kernel methods
- Boltzmann machines
- Tensor Networks
- Variational circuits
- Quantum Neural Networks
![Page 6: Nathan Killoran - Quantum Computer · 2019. 7. 31. · Nathan Killoran. Quantum computers are good at: Optimization Linear algebra Sampling Quantum physics. Quantum Machine Learning](https://reader033.vdocuments.net/reader033/viewer/2022051913/60043dc775bf5a42c606782e/html5/thumbnails/6.jpg)
LESSONS FROM DEEP LEARNING
![Page 7: Nathan Killoran - Quantum Computer · 2019. 7. 31. · Nathan Killoran. Quantum computers are good at: Optimization Linear algebra Sampling Quantum physics. Quantum Machine Learning](https://reader033.vdocuments.net/reader033/viewer/2022051913/60043dc775bf5a42c606782e/html5/thumbnails/7.jpg)
●
● Workhorse algorithms (backpropagation, stochasticgradient descent)
● Specialized, user-friendly software
● Hardware advancements (GPUs)
Why is Deep Learning successful?
![Page 8: Nathan Killoran - Quantum Computer · 2019. 7. 31. · Nathan Killoran. Quantum computers are good at: Optimization Linear algebra Sampling Quantum physics. Quantum Machine Learning](https://reader033.vdocuments.net/reader033/viewer/2022051913/60043dc775bf5a42c606782e/html5/thumbnails/8.jpg)
● Specialized, user-friendly software
● Hardware advancements (GPUs + QPUs)●
● Workhorse algorithms (quantum-aware backpropagation, stochastic gradient descent)
What can we leverage?
![Page 9: Nathan Killoran - Quantum Computer · 2019. 7. 31. · Nathan Killoran. Quantum computers are good at: Optimization Linear algebra Sampling Quantum physics. Quantum Machine Learning](https://reader033.vdocuments.net/reader033/viewer/2022051913/60043dc775bf5a42c606782e/html5/thumbnails/9.jpg)
TRAINING QUANTUM CIRCUITS
![Page 10: Nathan Killoran - Quantum Computer · 2019. 7. 31. · Nathan Killoran. Quantum computers are good at: Optimization Linear algebra Sampling Quantum physics. Quantum Machine Learning](https://reader033.vdocuments.net/reader033/viewer/2022051913/60043dc775bf5a42c606782e/html5/thumbnails/10.jpg)
Key Concepts for QML
● Variational circuits
● Quantum nodes
● Hybrid computation
● Quantum circuit learning
![Page 11: Nathan Killoran - Quantum Computer · 2019. 7. 31. · Nathan Killoran. Quantum computers are good at: Optimization Linear algebra Sampling Quantum physics. Quantum Machine Learning](https://reader033.vdocuments.net/reader033/viewer/2022051913/60043dc775bf5a42c606782e/html5/thumbnails/11.jpg)
Variational Circuits
I. Preparation of a fixed initial state
II. Quantum circuit; input data and free parameters are used as gate arguments
III. Measurement of fixed observable
● Main QML method for near-term (NISQ) devices
● Same basic structure as other modern algorithms:○ Variational Quantum Eigensolver (VQE)○ Quantum Alternating Operator Ansatz
(QAOA)
![Page 12: Nathan Killoran - Quantum Computer · 2019. 7. 31. · Nathan Killoran. Quantum computers are good at: Optimization Linear algebra Sampling Quantum physics. Quantum Machine Learning](https://reader033.vdocuments.net/reader033/viewer/2022051913/60043dc775bf5a42c606782e/html5/thumbnails/12.jpg)
II. Hardware-based- No access to quantum information; only have
measurements & expectation values- Needs to work as hardware becomes more powerful
and cannot be simulated
I. Simulator-based- Build simulation inside existing classical library- Can leverage existing optimization & ML tools- Great for small circuits, but not scalable
How to ‘train’ quantum circuits?
Two approaches:
![Page 13: Nathan Killoran - Quantum Computer · 2019. 7. 31. · Nathan Killoran. Quantum computers are good at: Optimization Linear algebra Sampling Quantum physics. Quantum Machine Learning](https://reader033.vdocuments.net/reader033/viewer/2022051913/60043dc775bf5a42c606782e/html5/thumbnails/13.jpg)
Gradients of quantum circuits
● Training strategy: use gradient descent algorithms.
● Need to compute gradients of variational circuit outputs w.r.t. their free parameters.
● How can we compute gradients of quantum circuits when even simulating their output is classically intractable?
![Page 14: Nathan Killoran - Quantum Computer · 2019. 7. 31. · Nathan Killoran. Quantum computers are good at: Optimization Linear algebra Sampling Quantum physics. Quantum Machine Learning](https://reader033.vdocuments.net/reader033/viewer/2022051913/60043dc775bf5a42c606782e/html5/thumbnails/14.jpg)
The ‘parameter shift’ trick
![Page 15: Nathan Killoran - Quantum Computer · 2019. 7. 31. · Nathan Killoran. Quantum computers are good at: Optimization Linear algebra Sampling Quantum physics. Quantum Machine Learning](https://reader033.vdocuments.net/reader033/viewer/2022051913/60043dc775bf5a42c606782e/html5/thumbnails/15.jpg)
Quantum Circuit Learning● Use the same device to compute a function and its gradient
○ Minimal overhead to compute gradients vs. original circuit
○ Optimize circuits using gradient descent
○ “Parameter shift” differentiation rule: gives exact gradients
○ Compatible with classical backpropagation: hybrid models are end-to-end differentiable
![Page 16: Nathan Killoran - Quantum Computer · 2019. 7. 31. · Nathan Killoran. Quantum computers are good at: Optimization Linear algebra Sampling Quantum physics. Quantum Machine Learning](https://reader033.vdocuments.net/reader033/viewer/2022051913/60043dc775bf5a42c606782e/html5/thumbnails/16.jpg)
Note:This is not finite differences!
• Exact• No restriction on the shift – in general,
we want a macroscopic shift
![Page 17: Nathan Killoran - Quantum Computer · 2019. 7. 31. · Nathan Killoran. Quantum computers are good at: Optimization Linear algebra Sampling Quantum physics. Quantum Machine Learning](https://reader033.vdocuments.net/reader033/viewer/2022051913/60043dc775bf5a42c606782e/html5/thumbnails/17.jpg)
![Page 18: Nathan Killoran - Quantum Computer · 2019. 7. 31. · Nathan Killoran. Quantum computers are good at: Optimization Linear algebra Sampling Quantum physics. Quantum Machine Learning](https://reader033.vdocuments.net/reader033/viewer/2022051913/60043dc775bf5a42c606782e/html5/thumbnails/18.jpg)
![Page 19: Nathan Killoran - Quantum Computer · 2019. 7. 31. · Nathan Killoran. Quantum computers are good at: Optimization Linear algebra Sampling Quantum physics. Quantum Machine Learning](https://reader033.vdocuments.net/reader033/viewer/2022051913/60043dc775bf5a42c606782e/html5/thumbnails/19.jpg)
Quantum Nodes
● Classical and quantum information are distinct
● QNode: common interface for quantum and classical devices
○ Classical device sees a callable parameterized function
○ Quantum device sees fine-grained circuit details
![Page 20: Nathan Killoran - Quantum Computer · 2019. 7. 31. · Nathan Killoran. Quantum computers are good at: Optimization Linear algebra Sampling Quantum physics. Quantum Machine Learning](https://reader033.vdocuments.net/reader033/viewer/2022051913/60043dc775bf5a42c606782e/html5/thumbnails/20.jpg)
Hybrid Computation● Use QPU with classical coprocessor
○ Classical optimization loop
○ Pre-/post-process quantum circuit outputs
○ Arbitrarily structured hybrid computations
![Page 21: Nathan Killoran - Quantum Computer · 2019. 7. 31. · Nathan Killoran. Quantum computers are good at: Optimization Linear algebra Sampling Quantum physics. Quantum Machine Learning](https://reader033.vdocuments.net/reader033/viewer/2022051913/60043dc775bf5a42c606782e/html5/thumbnails/21.jpg)
PENNYLANE
![Page 22: Nathan Killoran - Quantum Computer · 2019. 7. 31. · Nathan Killoran. Quantum computers are good at: Optimization Linear algebra Sampling Quantum physics. Quantum Machine Learning](https://reader033.vdocuments.net/reader033/viewer/2022051913/60043dc775bf5a42c606782e/html5/thumbnails/22.jpg)
• Train a quantum computerthe same way as a neural network
• Designed to scale as quantum computers grow in power
• Compatible with Xanadu, IBM, Rigetti, and Microsoft platforms https://github.com/XanaduAI/pennylane
https://pennylane.ai
PennyLane“The TensorFlow of quantum computing”
![Page 23: Nathan Killoran - Quantum Computer · 2019. 7. 31. · Nathan Killoran. Quantum computers are good at: Optimization Linear algebra Sampling Quantum physics. Quantum Machine Learning](https://reader033.vdocuments.net/reader033/viewer/2022051913/60043dc775bf5a42c606782e/html5/thumbnails/23.jpg)
Comes with a growing plugin ecosystem, supporting a wide range of quantum hardware and classical software
Q#
![Page 24: Nathan Killoran - Quantum Computer · 2019. 7. 31. · Nathan Killoran. Quantum computers are good at: Optimization Linear algebra Sampling Quantum physics. Quantum Machine Learning](https://reader033.vdocuments.net/reader033/viewer/2022051913/60043dc775bf5a42c606782e/html5/thumbnails/24.jpg)
PennyLane Example
![Page 25: Nathan Killoran - Quantum Computer · 2019. 7. 31. · Nathan Killoran. Quantum computers are good at: Optimization Linear algebra Sampling Quantum physics. Quantum Machine Learning](https://reader033.vdocuments.net/reader033/viewer/2022051913/60043dc775bf5a42c606782e/html5/thumbnails/25.jpg)
PennyLane Example
![Page 26: Nathan Killoran - Quantum Computer · 2019. 7. 31. · Nathan Killoran. Quantum computers are good at: Optimization Linear algebra Sampling Quantum physics. Quantum Machine Learning](https://reader033.vdocuments.net/reader033/viewer/2022051913/60043dc775bf5a42c606782e/html5/thumbnails/26.jpg)
• Run and optimize directly on quantumhardware (GPU→QPU)
• “Quantum-aware” implementation of backpropagation
• Hardware agnostic and extensible via plugins
• Open-source and extensively documented
• Use-cases:
○ Machine learning on large-scale quantum computations
○ Hybrid quantum-classical machinelearning
PennyLane Summary
https://github.com/XanaduAI/pennylane
https://pennylane.ai
![Page 27: Nathan Killoran - Quantum Computer · 2019. 7. 31. · Nathan Killoran. Quantum computers are good at: Optimization Linear algebra Sampling Quantum physics. Quantum Machine Learning](https://reader033.vdocuments.net/reader033/viewer/2022051913/60043dc775bf5a42c606782e/html5/thumbnails/27.jpg)