d-wavequantum computingaccess & applications via cloud deployment
TRANSCRIPT
Dr. Colin P. WilliamsVice President of Strategy & Corporate DevelopmentD-Wave Systems Inc.
D-Wave Quantum ComputingAccess & applications via cloud deployment
2Copyright © D-Wave Systems Inc.
D-Wave’s Mission & Activities
• Mission– To solve the world’s hardest problems especially in the
areas of artificial intelligence and machine learning
• Core technologies– Superconducting annealing-based quantum computers– Hybrid quantum/classical algorithms & architectures
• Business model– Quantum computer system sales– Quantum computer cloud services– Quantum machine learning services
3Copyright © D-Wave Systems Inc.
What are Quantum Computers?
4Copyright © D-Wave Systems Inc.
What are Quantum Computers?
Computers that harness quantum physical effects not available to conventional computers
Quantum Processor
5Copyright © D-Wave Systems Inc.
Superposition
Entanglement
Quantum Tunneling
Which Quantum Effects are Used?
6Copyright © D-Wave Systems Inc.
Our Approach in Context
• Annealing (D-Wave, Google, IARPA)– Harnesses Nature’s ability to find low energy configurations via quantum tunneling– Resilient to noise / does not require long coherence times / MIT pedigree– Handles a wide range of important problems– Currently non-universal but could be made universal
• Gate Model (Google, IBM, Intel, Alibaba, Rigetti)– Most common approach / based on analogy with Boolean logic circuits– Very difficult to scale; requires massive qubit overhead for error correction
• Topological (Microsoft)– Like gate model but without need for error correction (in theory)– Needs exotic quasi-particle whose robustness is now in dispute
(Phys. Rev. Lett., 118, 046801, 26 Jan 2017)
7Copyright © D-Wave Systems Inc.
D-Wave’sQuantum Computer
8Copyright © D-Wave Systems Inc.
Current Product: D-Wave 2000QTM
9Copyright © D-Wave Systems Inc.
World’s Most Advanced Quantum Processor
10Copyright © D-Wave Systems Inc.
Superconducting yet made in a CMOS Foundry
$1B World Class Production Facility 2000-qubit Circuits at 129,000 JJs .25μm design rules ASML 193nm lithography 65nm
11Copyright © D-Wave Systems Inc.
Functional Quantum Computation Established• Papers show superposition, entanglement & co-tunneling
– Johnson et al., “Q. Annealing with Manufactured Spins,” Nature 473, 194-198, 12th May (2011).– T. Lanting et al., “Cotunneling in pairs of coupled flux qubits,” Phys. Rev. B 82, 060512(R) (2010).– T. Lanting et al., “Entanglement in a Q. Annealing Processor,” Phys. Rev. X 4, 021041 (2014).
• These quantum effects play a functional role in the computations– Boixo, et al., "Computational multiqubit tunneling in programmable quantum annealers," Nature Communications 7,
Article number: 10327, Published 07 January (2016).
• UCL/USC showed that none of the classical models so far proposed as explanations for the D-Wave machine are correct– Albash et al., “Consistency Tests of Classical and Quantum Models for a Quantum Annealer,” Phys. Rev. A 91, 042314,
Published 13 April (2015).
• USC & D-Wave showed q. annealing can occur successfully on timescales orders of magnitude longer than the coherence time– Albash et al., "Decoherence in adiabatic quantum computation," Phys. Rev. A 91, 062320 (2015).– N G Dickson et al. “Thermally assisted quantum annealing of a 16-qubit problem”, Nature Communications 4, Article
number: 1903, 21 May (2013).
12Copyright © D-Wave Systems Inc.
Is Quantum Computing Ready for Deployment?
http://www.fz-juelich.de/ias/jsc/EN/Research/ModellingSimulation/QIP/QTRL/_node.html
JülichSupercomputing Center created a framework for Quantum Technology Readiness Levels
D-Wave
quantum
annealer
IBM
Experimental
qubit devices
13Copyright © D-Wave Systems Inc.
The Potential
14Copyright © D-Wave Systems Inc.
Potential for Massive Speedups
Copyright © 2017 D-Wave Systems Inc. 14
• Google found D-Wave 2X was 100,000,000x faster than QMC and SA on a particular problem (their “quantumess” test)
• More competitive classical algorithms become ineffective as we move to higher connectivity chips
See “What is the Computational Value of Finite Range Tunneling?” arXiv:1512.02206v3
• Apparent parallel scaling of QMC & D-Wave is an artifact (see E. Andriyash et al. “Can QMC simulate QA?” arXiv:1703.09277)
15Copyright © D-Wave Systems Inc.
Potential for Better Performance vs. Power Scaling
Source: http://www.extremetech.com/computing/116561-the-death-of-cpu-scaling-from-one-core-to-many-and-why-were-still-stuck
Quantum Power, ~0.1 μW
Classical Power
ClassicalPerformance
Quantum Performance
See: http://www.dwavesys.com/sites/default/files/14-1005A_D_wp_Computational_Power_Consumption_and_Speedup.pdf
• Transistors (000s)• Clock Speed (MHz)• Power (W)• Perf/Clock (ILP)
Rainier
Vesuvius
Washington
16Copyright © D-Wave Systems Inc.
Potential for Faster & Better (Lower Energy) SamplingEn
ergi
es o
f Sa
mp
les
Ret
urn
ed
Quantum ClassicalFat Tree #1
ClassicalFat Tree #2
Classical Selby
ClassicalSA #1
ClassicalSA #2
Best sample so far
17Copyright © D-Wave Systems Inc.
Quantum Sampling Accelerates Learning
• Specify model parameters θtrue, draw exact Boltzmann samples from θtrue, and estimate θ from samples
• Compare efficacy of CD, PCD, and QA-seeded MCMC chains at estimating the true distribution
• Compare rate of learning of a fully visible probabilistic graphical model classically vs. quantumly
Goal Model to Learn
Procedure
Classical Learning via PCDClassical Learning via CD
Quantum Learning (learns true θ faster)
D. Korenkevych et al., “Benchmarking Quantum Hardware for Training of Fully Visible Boltzmann Machines,” arXiv:1611.04528
Result: Quantum Learns Faster
18Copyright © D-Wave Systems Inc.
Factoring
Finding Ramsey
Numbers
Diagnosis
Constraint Satisfaction
Monte Carlo
FinancialModeling
SAT Filters
Discrete Sampling
Complexity Scaling
Quantum ErrorCorrecting Codes
QuantumSimulation
Quantum Research
Discrete Optimization
RadiotherapyOptimization
Deep Learning
VariationalAutoencoders
BoltzmannMachines
Applications Customers Have Mapped to D-Wave
Optimization Sampling Science
Structured Prediction
Trading Trajectory
OptimizationDetecting
Market Instability
19Copyright © D-Wave Systems Inc.
Lowest Level Programming
Computational Problem to Solve
Equivalent Ising Problem
Run Quantum Annealing
Read Out a Spin-configuration having Low Energy
Embedded Ising Problem
Retain all Solutions (SAMPLING)
Retain Best Solution(OPTIMIZATION)
Store Spin Configuration & Repeat
20Copyright © D-Wave Systems Inc.
Programming Languages & Software Tools
• Programming Languages– Python, Matlab, C/C++
• Compilers– Map problem to Ising problem
• Embedders– Ising problem to processor graph
• Pre-processing– Determining any forced variables– Decomposing larger problems
• Post-processing (optional)– Seeding heuristic optimizers– Inline GPU & CPU computations
21Copyright © D-Wave Systems Inc.
Applicationsin Finance
22Copyright © D-Wave Systems Inc.
PROBLEM: Invest $K amongst N assets at T time steps so as to maximize expected returns subject to varying risk and transaction costs at each time step
APPROACH: Quantum Optimization via D-Wave
• Couch problem as a quadratic integer optimization problem
• Map integer constraints to QUBOs
• Minimize sum of QUBOs via quantum annealing
Returnsat each time step
Transaction Costs
Sum of holdings at each time step = K
Max allowed holding of each asset = K’
Risk: Σ = forecast covariance matrix; γ = risk aversion
Optimization: Solving the Optimal Trading Trajectory Problem Using a Quantum Annealer arXiv1508.06182
IMPACT:Finds optimal strategy subjectto realisticconstraints
G. Roseberg et al (1QBit), M. L. de Prado (Guggenheim Partners), P. Carr (Courant Inst.) & K. Wu (LBL)
23Copyright © D-Wave Systems Inc.
PROBLEM: Seek signature of impending market instability by detecting onset of anomalously correlated moves
APPROACH: Model market as a graph; nodes = assets; edge if correlation > c
• Continually re-compute largest clique / Sudden expansion in clique size signals market move
Optimization: Impending Market Instability
IMPACT: Signals imminent market instability
24Copyright © D-Wave Systems Inc.
Application inHealthcare
25Copyright © D-Wave Systems Inc.
• Represent a model as a bit string
• HPC: bit string -> model -> simulation -> evaluation -> “score”
• D-Wave: Learn from results, adjust the program QC runs & iterate
• So QC guesses solution / HPC scores it / Adjust QC & repeat
Quantum-Accelerated Optimization
Better “guesses”
“Scores” for the guesses
Use QC to reduce number of calls to HPC
26Copyright © D-Wave Systems Inc.
Why Does it Work?
• Imagine you’ve reached an intermediate point in design space and want to pick the next bit string to try
• Classical methods only sense the local neighborhood
• Quantum methods have potential for greater horizon
• Make a better next move possibly leading in different direction
Classical
Dis
crep
ancy
Design Parameter
Quantum
Dis
crep
ancy
Design Parameter
27Copyright © D-Wave Systems Inc.
Case Study: Radiotherapy Optimization
PROBLEM: Deliver lethal dose to tumor whilst minimizing damage to healthy tissues
APPROACH: Hybrid: QC + Conventional Computer• Radiation treatment plan = bit string• Quality = result of running extensive
radiation transport simulation• Results of radiation transport simulations
drive adjustments to plan
IMPACT:• Hybrid quantum-classical design found a radiation therapy
treatment that minimized the objective function to 70.7 c.f. 120.0 for tabu, and ran in 1/3 the time making fewer calls to radiation transport sim. Source: “Varian’s RapidArc Radiation Delivery
System Goes Clinical,” medGadget, July 22nd (2008)
28Copyright © D-Wave Systems Inc.
Applications inMachine Learning
29Copyright © D-Wave Systems Inc.
People Have Used D-Wave for Machine Learning
• Mainly supervised learning so far– Yes/No classifier for cars in images
– Wink/Blink classifier
– Yes/No Higgs boson event classifier
Source: A. Mott, J. Job, J. Vlimant, D. Lidar & M. Spiropulu, Nature, Volume 550, 19th October 2017.
30Copyright © D-Wave Systems Inc.
What Current A.I. Doesn’t Do Well
“Unsupervised learning had a catalytic effect in reviving interest indeep learning, but has since been overshadowed by the successes ofpurely supervised learning. […] we expect unsupervised learning tobecome far more important in the longer term. Human and animallearning is largely unsupervised: we discover the structure of theworld by observing it, not by being told the name of every object.”
Yann LeCun, Yoshua Bengio & Geoffrey Hinton,“Deep Learning,” Nature, Vol. 521, 28th May (2015)
• A.I. is not yet sufficiently good at unsupervised learning
31Copyright © D-Wave Systems Inc.
Quantum Computers Can Help
• Unsupervised learning can use probabilistic models
• These rely on sampling
• Quantum computers have potential to revolutionize A.I. by making unsupervised learning models feasible to train efficiently
• Because quantum computers are fast native samplers
32Copyright © D-Wave Systems Inc.
How does Sampling Arise in Probabilistic Models?
• Consider a Restricted Boltzmann Machine (RBM)
• RBMs can be components of more complex neural networks
Model parameters q º {bi}È{c j}È{Wi j}
Visible units with biases bi
Hidden units with biases cj
W1,1W1,2
W10,10
33Copyright © D-Wave Systems Inc.
What does Training Entail? Discrete Sampling!
• Given training data (visible vectors) vt s.t. 1 ≤ t ≤ T
• Adjust model parameters s.t. model most likely reproduces the training data
• Done by maximizing the log-likelihood of the observed data distribution w.r.t. the model parameters, θi
¶
¶qilog(p(vt ))
t=1
T
åæ
è
çç
ö
ø
÷÷= -
¶
¶qiE(vt,h)
t=1
T
åp(h |v
t)
+T¶
¶qiE(v, h)
p(v, h)
• Positive Phase
• Expectation over p(h|vt) in “clamped” condition
• Requires sampling over the (given) data distribution
• Simple!
• Negative Phase
• Expectation over p(v, h) in “unclamped” condition
• Requires sampling over the (predicted) model distribution
• Intractable!
where
34Copyright © D-Wave Systems Inc.
ooooDirected hierarchical
aanetwork of continuousvariables
Directed hierarchical
aanetwork of continuousoooovariables
Discrete Sampling in Complex Architectures (DVAE/QVAE)
• Real data has discrete & continuous variables
• Natural to want discrete hidden variables
• Can’t backpropagate through discrete variables
• DVAE solves this problem– See J. Rolfe, “Discrete Variational Autoencoders”,
arXiv:1609.02200 Undirected network of discrete variables (a fully hidden bipartite (possibly quantum) Boltzmann machine
Encoder Network
Decoder Network
Sample from smoothed distribution
Smoothed (i.e., continuous) discrete latent variables
Random samples from U[0,1]
• Exceeds state of the art on three standard machine learning datasets
• DVAE (classical) / QVAE (quantum)
35Copyright © D-Wave Systems Inc.
Training Digits
DVAE
Previous State of the Art
J. Rolfe, “Discrete Variational Autoencoders”, arXiv:1609.02200 [stat.ML]
DVAE Exceeds State-of-the-Art on a Generative Task
Machine-Generated Novel DigitsReconstructed Digits
36Copyright © D-Wave Systems Inc.
Why we Believe QVAE will Further Improve DVAE
State-of-the-Art (classical) DVAE (classical) QVAE (via QMC simulation)
Same number of training epochs used in each case
37Copyright © D-Wave Systems Inc.
Quantum/Classical Machine Learning Services
• D-Wave web services are designed to make it easier to train PML models
• Capabilities– Learns from noisy / incomplete data– Quantifies confidence in predictions– Reveals hidden correlations in data– Infers missing data
• Functionality (Web Services for PML):– Classical Boltzmann sampling (GPU)– Quantum Boltzmann sampling
(CPU)– Raw QPU sampling (QPU)
• Supports both ML/QML models
• Called from TensorFlow or Python
38Copyright © D-Wave Systems Inc.
HybridQuantum/Classical
Cloud Model
39Copyright © D-Wave Systems Inc.
Quantum Cloud
• Quantum computers …– Need typical datacenter infrastructure
– Are expensive
– Are evolving fast / quickly obsolete
– Are alien to most programmers
• Solution– Make QCs accessible via the cloud
– Bill usage by minute or by time blocks
– Lease QC time rather than buy QC
– Access via familiar software & tools
– Includes classical proxy solvers for code development & debugging
40Copyright © D-Wave Systems Inc.
Bill Gates Expects Quantum Cloud Access [by 2026]
Source: http://www.zdnet.com/article/quantum-cloud-computing-could-arrive-in-the-next-decade-says-bill-gates/
41Copyright © D-Wave Systems Inc.
D-Wave Sells Quantum Cloud Access Already
Jane Edwards, "D-Wave to Provide Oak Ridge National Lab Cloud Access to Quantum Computing", ExecutiveBiz.com blog post, July 26th, (2017) web URLhttp://blog.executivebiz.com/2017/07/d-wave-to-provide-oak-ridge-national-lab-cloud-access-to-quantum-computing-platform/
42Copyright © D-Wave Systems Inc.
Conclusions
• Quantum computing will turbo-charge unsupervised learning
• Quantum and hybrid machine learning models already running
• Our first web-ML services were released in September 2017
– Reinvigorate probabilistic machine learning and prepare ground for future quantum & hybrid ML services
– Both state-of-the-art today (classically) / faster tomorrow (quantumly)
• DVAE already surpassing state of the art / QVAE coming in 2018
• Seeking users for our quantum cloud services!
Contact: [email protected]
43Copyright © D-Wave Systems Inc.
Thank you!Email: [email protected]