[QIP-Sem] QIP seminar, Mon 2/22, 4:30, 36-428, Denchev,Vasil

[Peter Shor]

MIT Quantum Information Processing seminar
Monday 2/22 at 4:30 in 36-428
-------------------------------------------------

 Denchev,Vasil (Purdue/Google)

Applying Adiabatic Quantum Computing to Machine Learning

Abstract:

We report on work aiming to apply adiabatic quantum computing (AQC) to machine learning. Machine learning is of rapidly growing importance to many engineering disciplines. At their core, most machine learning problems map to formally NP-hard optimization. This motivates the use of AQC as it promises to deliver solutions of higher quality than classical heuristic solvers. Specifically, we study the problem of training a binary classifier. The classifier is constructed as a thresholded linear superposition of a set of weak classifiers with weights that are optimized in a learning process striving to minimize the training error as well as the number of weak classifiers used. We show that the necessary bit precision of the weights grows only logarithmically with the ratio of the number of training examples to the number of weak classifiers and formulate the training as quadratic unconstrained binary optimization to conform to the format required by D-Wave's implementation of AQC.
  Next, we generalize this baseline method to large-scale classifier training, where either the cardinality of the dictionary of weak classifiers or the number of weak classifiers used in the final strong classifier exceed the number of variables that can be handled in a single global optimization. For such situations we propose an iterative and piecewise approach in which a subset of weak classifiers is selected in each iteration via global optimization. The strong classifier is then constructed by concatenating the subsets of weak classifiers. We find that on our benchmark problems the proposed method outperforms the state-of-the-art algorithm AdaBoost, even when we use a classical heuristic as a stand-in for the quantum hardware. We also provide theoretical arguments as to why the proposed optimization method, which does not only minimize the training error but also adds L0-norm regularization, is superior to versions of boosting that only minimize the training error. By c
 onducting a quantum Monte Carlo simulation, we gather initial evidence that AQC may be able to handle a generic training problem efficiently. Finally, we describe our first attempt to use D-Wave's current hardware to train a large-scale detector.

-------------------------------------------------
http://qis.mit.edu
http://mailman.mit.edu/mailman/listinfo/qip-sem