sample-optimal learning of graphical models · learning discrete graphical models with generalized...
TRANSCRIPT
Los Alamos National Laboratory
Sample-Optimal Learning of Graphical Models(and opening the D-Wave annealer’s quantum box)
Andrey Lokhovwith M. Chertkov, C. Coffrin, Y. Kharkov, S. Misra, M. Vuffray
Theoretical Division, Los Alamos National Laboratory
Learning General Markov Random Fields Performance and Rigorous Guarantees
Quantum Annealing and Inverse Ising Problem Understanding the Behavior of D-Wave Annealer
𝜎𝑖 𝜎𝑗𝐽𝑖𝑗𝜎𝑖𝜎𝑗
Input Output
𝐽𝑖𝑗Couplings
Fields ℎ𝑖 𝜎
Annealing
Spin configurations
𝐽1
𝐽2
𝐽𝑠Indistinguishable
Distributions
𝐽1
𝐽2
𝜉
𝜉
𝜉
)( kkf
Arbitrary alphabet
Arbitrary interaction orders
qAA iii ,
Lk
Arbitrary basis functions )( kkf
iKk
ki
k
Kk
kkk fZ
***max,min))(exp
1)(
Example: binary variables and multi-body interactions
Learning problem: given M i.i.d. samples , with probability learn the graph of conditional dependencies and estimate couplings
Regularized Interaction Screening Estimator (RISE): convex optimization
Optimal Structure and Parameter Learning of Ising Models
A. Lokhov, M. Vuffray, S. Misra, M. Chertkov
Interaction Screening: Efficient and Sample-Optimal Learning of Ising Models
M. Vuffray, S. Misra, A. Lokhov, M. Chertkov
(2018)
(2016)
Theorem : If , then with prob.
Our Method: Interaction Screening
kj
ijk
iijkj
ij
iij
i
ii JJh ***
exp)(
1
,
expminargˆ,ˆ ikj
jk
iijkj
j
iijiihJ
ii JJJhhJii
NqCM
L
ln)6exp(
2
2
2
ˆ2
* ii 1
i
1 )()1(,,
M
*
Learning Discrete Graphical Models with Generalized Interaction Screening
S. Misra, A. Lokhov, M. Vuffray (submitted)
Exact, tractable & parallelizable algorithm
Requires near IT optimal number of samples
Works for any models: low temperature, spin glass, …
Comparison with IT bounds for Ising models:
Generalizations: dense models, reconstruction in the dynamic setting, …
+ assumptions-free proof for logistic regression loss (pseudo-likelihood)
Classical annealing Quantum annealing
Tem
per
atu
re
Ob
ject
ive
zx sHHssH )1()(
D-Wave computer is an implementation of quantum annealing for Ising models
Motivation: assessing performance and quantify uncertainties
Solving Inverse Ising problem
Question 2: What can we learn about the machine from this distribution?
• Type of interactions?• Structure of interactions? • Input-output response?
Question 1: Characterize the probability distribution of output data
Mixture of models withfluctuating magnetic fields
Single model with aspurious coupling
Statistical reconstruction of instantaneous field fluctuations spurious links with quadraticresponse
Spurious interactions arequadratic response to physicalinteractions forming “triangles”
Explanation for spurious links:
Structure of distribution:
Input-output response:
Opening the Quantum Box: Understanding the Behavior of Analogue Annealers
with Statistical Learning A. Lokhov, Y. Kharkov, C. Coffrin, M. Vuffray (submitted)
56 Three-Body Terms