stanislav fort with neural net works & differentiable...
TRANSCRIPT
Learning Adaptive Quantum State Tomography with
Neural Networks & Differentiable Programming
Stanislav FortStanford University (prev. Google Research)
[email protected]/~sfort1/
@stanislavfort
Stanislav Fort | Adaptive Quantum State Tomography with NNs and differentiable programming | 19 February 2020 | Stanford, USA
Primarily based on arXiv 1812.06693
Primarily based on arXiv 1812.06693, in review
Stanislav Fort | Adaptive Quantum State Tomography with NNs and differentiable programming | 19 February 2020 | Stanford, USA
Stanislav Fort | Adaptive Quantum State Tomography with NNs and differentiable programming | 19 February 2020 | Stanford, USA
Primarily based on arXiv 1812.06693, in review
AI, ML, DL & differentiable programming
Stanislav Fort | Adaptive Quantum State Tomography with NNs and differentiable programming | 19 February 2020 | Stanford, USA
● Artificial intelligence: The science of making machines smart
● Machine learning: Machines getting smart from data
● Deep learning: Machines getting smart from data using deep neural networks as functional approximators
● Differentiable programming:Taking partial derivatives through programs, not restricted to deep neural networks as functions
AI, ML, DL & differentiable programming
Stanislav Fort | Adaptive Quantum State Tomography with NNs and differentiable programming | 19 February 2020 | Stanford, USA
● Artificial intelligence: The science of making machines smart
● Machine learning: Machines getting smart from data
● Deep learning: Machines getting smart from data using deep neural networks as functional approximators
● Differentiable programming:Taking partial derivatives through programs, not restricted to deep neural networks as functions
AI, ML, DL & differentiable programming
Stanislav Fort | Adaptive Quantum State Tomography with NNs and differentiable programming | 19 February 2020 | Stanford, USA
● Artificial intelligence: The science of making machines smart
● Machine learning: Machines getting smart from data
● Deep learning: Machines getting smart from data using deep neural networks as functional approximators
● Differentiable programming:Taking partial derivatives through programs, not restricted to deep neural networks as functions
AI, ML, DL & differentiable programming
Stanislav Fort | Adaptive Quantum State Tomography with NNs and differentiable programming | 19 February 2020 | Stanford, USA
● Artificial intelligence: The science of making machines smart
● Machine learning: Machines getting smart from data
● Deep learning: Machines getting smart from data using deep neural networks as functional approximators
● Differentiable programming:Taking partial derivatives through programs, not restricted to deep neural networks as functions
Differentiable programmingHeat equation example
Stanislav Fort | Adaptive Quantum State Tomography with NNs and differentiable programming | 19 February 2020 | Stanford, USA
∂
∂ temperature field
Bridging ML &scientific computing
Differentiable programmingQuantum games example
Stanislav Fort | Adaptive Quantum State Tomography with NNs and differentiable programming | 19 February 2020 | Stanford, USA
∂
∂ quantum game strategy
Writing down a solution vs learning a solution
Stanislav Fort | Adaptive Quantum State Tomography with NNs and differentiable programming | 19 February 2020 | Stanford, USA
f( )= “tortoise”
Bubble sort (explicit)
Image classification(learned)
Translation(much better when learned)
Writing down a solution vs learning a solution
Stanislav Fort | Adaptive Quantum State Tomography with NNs and differentiable programming | 19 February 2020 | Stanford, USA
Even solving symbolic math
Machine learningLearning to get faster (& better) heuristics
Stanislav Fort | Adaptive Quantum State Tomography with NNs and differentiable programming | 19 February 2020 | Stanford, USA
Machine learningLearning to get faster (& better) heuristics
Stanislav Fort | Adaptive Quantum State Tomography with NNs and differentiable programming | 19 February 2020 | Stanford, USA
Network & data structureHow to induce the correct learning prior?
Stanislav Fort | Adaptive Quantum State Tomography with NNs and differentiable programming | 19 February 2020 | Stanford, USA
Stanislav Fort | Adaptive Quantum State Tomography with NNs and differentiable programming | 19 February 2020 | Stanford, USA
Quantum state tomography
Stanislav Fort | Adaptive Quantum State Tomography with NNs and differentiable programming | 19 February 2020 | Stanford, USA
Quantum state tomographyPure states first
Stanislav Fort | Adaptive Quantum State Tomography with NNs and differentiable programming | 19 February 2020 | Stanford, USA
Quantum state tomographyPure states first
Stanislav Fort | Adaptive Quantum State Tomography with NNs and differentiable programming | 19 February 2020 | Stanford, USA
Quantum state tomographyPure states first
Stanislav Fort | Adaptive Quantum State Tomography with NNs and differentiable programming | 19 February 2020 | Stanford, USA
Quantum state tomographyPure states first
Stanislav Fort | Adaptive Quantum State Tomography with NNs and differentiable programming | 19 February 2020 | Stanford, USA
Quantum state tomographyPure states first
Stanislav Fort | Adaptive Quantum State Tomography with NNs and differentiable programming | 19 February 2020 | Stanford, USA
Quantum state tomographyPure states first
? ??
??
Stanislav Fort | Adaptive Quantum State Tomography with NNs and differentiable programming | 19 February 2020 | Stanford, USA
Quantum state tomographyPure states first
Stanislav Fort | Adaptive Quantum State Tomography with NNs and differentiable programming | 19 February 2020 | Stanford, USA
Quantum state tomographyPure states first
? ??
??
Stanislav Fort | Adaptive Quantum State Tomography with NNs and differentiable programming | 19 February 2020 | Stanford, USA
Quantum state tomographyPure states first
? ??
??
(Adaptivity)
Stanislav Fort | Adaptive Quantum State Tomography with NNs and differentiable programming | 19 February 2020 | Stanford, USA
Quantum state tomographyPure states first
? ????
Stanislav Fort | Adaptive Quantum State Tomography with NNs and differentiable programming | 19 February 2020 | Stanford, USA
Quantum state tomographyPure states first
? ??????
???
Stanislav Fort | Adaptive Quantum State Tomography with NNs and differentiable programming | 19 February 2020 | Stanford, USA
Quantum state tomographyPure states first
? ??????
?????
???
Stanislav Fort | Adaptive Quantum State Tomography with NNs and differentiable programming | 19 February 2020 | Stanford, USA
Quantum state tomographyPure states first
? ??????
?????
?????
???
N
Stanislav Fort | Adaptive Quantum State Tomography with NNs and differentiable programming | 19 February 2020 | Stanford, USA
Quantum state tomographyPure states first
? ????? ????? ????? ??
??
N
Stanislav Fort | Adaptive Quantum State Tomography with NNs and differentiable programming | 19 February 2020 | Stanford, USA
Quantum state tomographyPure states first
? ????? ????? ????? ??
??
N
Stanislav Fort | Adaptive Quantum State Tomography with NNs and differentiable programming | 19 February 2020 | Stanford, USA
Quantum state tomographyWhat is difficult?
? ????? ????? ????? ??
??
N
Stanislav Fort | Adaptive Quantum State Tomography with NNs and differentiable programming | 19 February 2020 | Stanford, USA
Pure states → density matricesDensity matricesPure states
Stanislav Fort | Adaptive Quantum State Tomography with NNs and differentiable programming | 19 February 2020 | Stanford, USA
What do we care about?
1) How precisely can you reconstruct the unknown state, given N copies of the unknown state?
2) How much compute does it take?
Stanislav Fort | Adaptive Quantum State Tomography with NNs and differentiable programming | 19 February 2020 | Stanford, USA
Regular Quantum State Tomography
Stanislav Fort | Adaptive Quantum State Tomography with NNs and differentiable programming | 19 February 2020 | Stanford, USA
Regular Quantum State TomographyStage 1: Take measurements
Stanislav Fort | Adaptive Quantum State Tomography with NNs and differentiable programming | 19 February 2020 | Stanford, USA
Regular Quantum State TomographyStage 2: Process them
Stanislav Fort | Adaptive Quantum State Tomography with NNs and differentiable programming | 19 February 2020 | Stanford, USA
Regular Quantum State TomographyProblem: Measurements expensive
Stanislav Fort | Adaptive Quantum State Tomography with NNs and differentiable programming | 19 February 2020 | Stanford, USA
Adaptive Quantum State TomographyFerenc Huszár, Neil M. T. Houlsby. Adaptive Bayesian Quantum Tomography, arXiv 1107.0895
Stage 1: Take measurements up to t
Stanislav Fort | Adaptive Quantum State Tomography with NNs and differentiable programming | 19 February 2020 | Stanford, USA
Adaptive Quantum State TomographyFerenc Huszár, Neil M. T. Houlsby. Adaptive Bayesian Quantum Tomography, arXiv 1107.0895
Stage 2: Determine the next POVM
Stanislav Fort | Adaptive Quantum State Tomography with NNs and differentiable programming | 19 February 2020 | Stanford, USA
Adaptive Quantum State TomographyParticle bank maintaining and update
Stanislav Fort | Adaptive Quantum State Tomography with NNs and differentiable programming | 19 February 2020 | Stanford, USA
Adaptive Quantum State TomographyParticle bank maintaining and update
Stanislav Fort | Adaptive Quantum State Tomography with NNs and differentiable programming | 19 February 2020 | Stanford, USA
Adaptive Quantum State TomographyParticle bank resampling
Stanislav Fort | Adaptive Quantum State Tomography with NNs and differentiable programming | 19 February 2020 | Stanford, USA
Adaptive Quantum State Tomography
Problem: Computationally expensive
Stanislav Fort | Adaptive Quantum State Tomography with NNs and differentiable programming | 19 February 2020 | Stanford, USA
Neural Adaptive Quantum State Tomography (NA-QST)
Our approachYihui Quek*, Stanislav Fort*, Hui Khoon Ng. Adaptive Quantum State Tomography with Neural Networks, arXiv 1812.06693
Advantages:- Parametrized state ansatz is not
required- Exponential computational speedup- Learned directly from simulated data- Can retrain within hours- Any POVM types- Different noise models- Learned distance measure, update rule
etc.
Stanislav Fort | Adaptive Quantum State Tomography with NNs and differentiable programming | 19 February 2020 | Stanford, USA
Neural Adaptive Quantum State Tomography (NA-QST)Custom recurrent architecture - off the shelf doesn’t work
Stanislav Fort | Adaptive Quantum State Tomography with NNs and differentiable programming | 19 February 2020 | Stanford, USA
Neural Adaptive Quantum State Tomography (NA-QST)Custom recurrent architecture - off the shelf doesn’t work
Train: Differentiable quantum simulator provides measurement results
Stanislav Fort | Adaptive Quantum State Tomography with NNs and differentiable programming | 19 February 2020 | Stanford, USA
Neural Adaptive Quantum State Tomography (NA-QST)Custom recurrent architecture - off the shelf doesn’t work
Train: Differentiable quantum simulator provides measurement results
Test: Experimenter provides measurement results
Stanislav Fort | Adaptive Quantum State Tomography with NNs and differentiable programming | 19 February 2020 | Stanford, USA
Neural Adaptive Quantum State Tomography (NA-QST)Inside the RNN cell - where the magic happens
Stanislav Fort | Adaptive Quantum State Tomography with NNs and differentiable programming | 19 February 2020 | Stanford, USA
Neural Adaptive Quantum State Tomography (NA-QST)Inside the RNN cell - where the magic happens
Stanislav Fort | Adaptive Quantum State Tomography with NNs and differentiable programming | 19 February 2020 | Stanford, USA
Neural Adaptive Quantum State Tomography (NA-QST)Inside the RNN cell - where the magic happens
Stanislav Fort | Adaptive Quantum State Tomography with NNs and differentiable programming | 19 February 2020 | Stanford, USA
Neural Adaptive Quantum State Tomography (NA-QST)Inside the RNN cell - where the magic happens
Stanislav Fort | Adaptive Quantum State Tomography with NNs and differentiable programming | 19 February 2020 | Stanford, USA
Neural Adaptive Quantum State Tomography (NA-QST)Inside the RNN cell - where the magic happens
Stanislav Fort | Adaptive Quantum State Tomography with NNs and differentiable programming | 19 February 2020 | Stanford, USA
Neural Adaptive Quantum State Tomography (NA-QST)Inside the RNN cell - where the magic happens
Veil of ignorance =
gradient stop
Stanislav Fort | Adaptive Quantum State Tomography with NNs and differentiable programming | 19 February 2020 | Stanford, USA
Neural Adaptive Quantum State Tomography (NA-QST)Inside the RNN cell - where the magic happens
Stanislav Fort | Adaptive Quantum State Tomography with NNs and differentiable programming | 19 February 2020 | Stanford, USA
Neural Adaptive Quantum State Tomography (NA-QST)Inside the RNN cell - where the magic happens
Stanislav Fort | Adaptive Quantum State Tomography with NNs and differentiable programming | 19 February 2020 | Stanford, USA
Neural Adaptive Quantum State Tomography (NA-QST)Inside the RNN cell - where the magic happens
Stanislav Fort | Adaptive Quantum State Tomography with NNs and differentiable programming | 19 February 2020 | Stanford, USA
Neural Adaptive Quantum State Tomography (NA-QST)Inside the RNN cell - where the magic happens
Stanislav Fort | Adaptive Quantum State Tomography with NNs and differentiable programming | 19 February 2020 | Stanford, USA
Neural Adaptive Quantum State Tomography (NA-QST)Inside the RNN cell - where the magic happens
Stanislav Fort | Adaptive Quantum State Tomography with NNs and differentiable programming | 19 February 2020 | Stanford, USA
Neural Adaptive Quantum State Tomography (NA-QST)Inside the RNN cell - where the magic happens
Stanislav Fort | Adaptive Quantum State Tomography with NNs and differentiable programming | 19 February 2020 | Stanford, USA
Neural Adaptive Quantum State Tomography (NA-QST)Inside the RNN cell - where the magic happens
Stanislav Fort | Adaptive Quantum State Tomography with NNs and differentiable programming | 19 February 2020 | Stanford, USA
Neural Adaptive Quantum State Tomography (NA-QST)Inside the RNN cell - where the magic happens
Stanislav Fort | Adaptive Quantum State Tomography with NNs and differentiable programming | 19 February 2020 | Stanford, USA
Neural Adaptive Quantum State Tomography (NA-QST)Inside the RNN cell - where the magic happens
Stanislav Fort | Adaptive Quantum State Tomography with NNs and differentiable programming | 19 February 2020 | Stanford, USA
Neural Adaptive Quantum State Tomography (NA-QST)Inside the RNN cell - where the magic happens
Stanislav Fort | Adaptive Quantum State Tomography with NNs and differentiable programming | 19 February 2020 | Stanford, USA
Neural Adaptive Quantum State Tomography (NA-QST)Different paradigm from ABQT
● No explicit Bayesian interpretation - weights are just weights
● Automatically learned NN similarity metric for resampling and weight updates
● Resampling is fast ● End-to-end training minimizing arbitrary human
choices
Stanislav Fort | Adaptive Quantum State Tomography with NNs and differentiable programming | 19 February 2020 | Stanford, USA
Neural Adaptive Quantum State Tomography (NA-QST)How is it trained? 2 notions of “time”
Inference step
Stanislav Fort | Adaptive Quantum State Tomography with NNs and differentiable programming | 19 February 2020 | Stanford, USA
Neural Adaptive Quantum State Tomography (NA-QST)How is it trained? 2 notions of “time”
Backprop step
Stanislav Fort | Adaptive Quantum State Tomography with NNs and differentiable programming | 19 February 2020 | Stanford, USA
Neural Adaptive Quantum State Tomography (NA-QST)Numerical experiments - what can be varied?
1) Single-qubit POVM type: 2 (basis POVM), 3, 4 (SIC), and 6 (Pauli) legs per qubit subspace
Stanislav Fort | Adaptive Quantum State Tomography with NNs and differentiable programming | 19 February 2020 | Stanford, USA
Neural Adaptive Quantum State Tomography (NA-QST)Numerical experiments - what can be varied?
1) Single-qubit POVM type: 2 (basis POVM), 3, 4 (SIC), and 6 (Pauli) legs per qubit subspace
2) POVM Adaptivity: Adaptive or random measurements? What are the benefits?
3) 3) Reconstruction algorithm: Standard QST, ABQT, or our NA-QST
Stanislav Fort | Adaptive Quantum State Tomography with NNs and differentiable programming | 19 February 2020 | Stanford, USA
Neural Adaptive Quantum State Tomography (NA-QST)Reconstruction accuracy and time to compute
Stanislav Fort | Adaptive Quantum State Tomography with NNs and differentiable programming | 19 February 2020 | Stanford, USA
Neural Adaptive Quantum State Tomography (NA-QST)Reconstruction accuracy and time to compute
Number of measurements
Stanislav Fort | Adaptive Quantum State Tomography with NNs and differentiable programming | 19 February 2020 | Stanford, USA
Neural Adaptive Quantum State Tomography (NA-QST)Reconstruction accuracy and time to compute
Reconstruction error
Stanislav Fort | Adaptive Quantum State Tomography with NNs and differentiable programming | 19 February 2020 | Stanford, USA
Neural Adaptive Quantum State Tomography (NA-QST)Reconstruction accuracy and time to compute
NA-QST is:
1) Equal in reconstruction accuracy
Stanislav Fort | Adaptive Quantum State Tomography with NNs and differentiable programming | 19 February 2020 | Stanford, USA
Neural Adaptive Quantum State Tomography (NA-QST)Reconstruction accuracy and time to compute
NA-QST is:
1) Equal in reconstruction accuracy
2) Orders of magnitude faster!
Stanislav Fort | Adaptive Quantum State Tomography with NNs and differentiable programming | 19 February 2020 | Stanford, USA
Neural Adaptive Quantum State Tomography (NA-QST)Reconstruction accuracy and time to compute
NA-QST is:
1) Equal in reconstruction accuracy
2) Orders of magnitude faster!
3) Time complexity scales better
Stanislav Fort | Adaptive Quantum State Tomography with NNs and differentiable programming | 19 February 2020 | Stanford, USA
Neural Adaptive Quantum State Tomography (NA-QST)Time complexity scaling: polynomial to logarithmic
NA-QST: logarithmic time scaling
ABQT: polynomial time scaling
Stanislav Fort | Adaptive Quantum State Tomography with NNs and differentiable programming | 19 February 2020 | Stanford, USA
Neural Adaptive Quantum State Tomography (NA-QST)When does adaptivity help?
Conclusions:
- 2 legs (basis): Adaptivity helps a lot
- 3 legs: Adaptive helps slightly
- 4 legs and above (informationally complete): Adaptivity does not make any difference
Stanislav Fort | Adaptive Quantum State Tomography with NNs and differentiable programming | 19 February 2020 | Stanford, USA
Takeaways and thank you!We designed, implemented, and tested an end-to-end trainable, deep learning powered algorithm called Neural Adaptive Quantum State Tomography (NA-QST)
NA-QST is:1) Fast to train (hours on a laptop)
2) Very fast to run (poly → log)3) Accurate in reconstruction (~SOTA)4) Flexible (noise, different POVMs etc)
Future: Retraining for downstream products involving the density matrix
Primarily based on arXiv 1812.06693