Totally corrective boosting with cardinality penalization
Methods, systems, and apparatus, for totally corrective boosting with cardinality penalization are described. One of the methods includes obtaining initialization data identifying training examples, a dictionary of weak classifiers, and an active weak classifier matrix. Iterations of a totally corrective boosting with cardinality penalization process are performed, wherein each iteration performs operations comprising selecting a weak classifier from the dictionary of weak classifiers that most violates a constraint of a dual of the primal problem. The selected weak classifier is included in the active weak classifier matrix. The primal problem is optimized, and a discrete weight vector is determined. Weak classifiers are identified from the active weak classifier matrix with respective discrete weights greater than a threshold. The regularized risk is optimized, and a continuous weight vector is determined. The classifier is determined as an ensemble identified by the weak classifiers and the continuous weight vector.
1. A system comprising:
a programmable quantum annealing chip and one or more computers, the programmable quantum annealing chip comprising a plurality of qubits, wherein the plurality of qubits are connected by programmable couplers, and one or more storage devices storing instructions that are operable, when executed by the programmable quantum annealing chip and the one or more computers, to cause the programmable quantum annealing chip and the one or more computers to perform operations comprising:
obtaining initialization data identifying a plurality of training examples and an active weak classifier matrix; and
performing one or more iterations of a totally corrective boosting with cardinality penalization process until a termination condition is satisfied, wherein performing the one or more iterations comprises:
selecting a weak classifier from a dictionary of weak classifiers,
including the selected weak classifier in the active weak classifier matrix,
performing quantum annealing with the quantum annealing chip to optimize a primal problem based, in part, on the training examples, the active weak classifier matrix, and a plurality of classifier weights, to select one or more weak classifiers with respective discrete weights greater than a threshold, wherein the classifier weights are mapped to the qubits of the quantum annealing chip, and
optimizing 1-regularized risk using the identified one or more weak classifiers and the training to provide a set of continuous classifier weights for the selected active weak classifiers, and
determining whether the termination condition is satisfied, wherein determining whether the termination condition is satisfied comprises determining whether a maximum number of iterations has been performed, wherein when the maximum number of iterations has been performed, the process comprises storing in memory the active weak classifier matrix and the set of continuous classifier weights, and wherein when the maximum number of iterations has not been performed, performing another iteration of the totally corrective boosting with cardinality penalization process.
2. The system of claim 1 , wherein quantum annealing is performed on an Ising model or a restricted Ising model.
3. The system of claim 1 , wherein the plurality of qubits comprise superconducting loops.
4. The system of claim 1 , wherein each unit cell comprises 8 qubits.
5. The system of claim 1 , wherein the programmable couplers comprise inductive couplers.
6. The system of claim 1 , wherein optimizing the primal problem includes using, wherein each of the discrete weights is a non negative fixed point n-bit value, and each bit of each of the n-bit values is mapped to a corresponding quantum bit of the programmable quantum annealing chip.
7. A quantum-computer implemented method comprising:
obtaining, by a programmable quantum annealing chip, initialization data identifying a plurality of training examples and an active weak classifier matrix, wherein the programmable quantum annealing chip comprises a plurality of qubits connected by programmable couplers;
performing one or more iterations of a totally corrective boosting with cardinality penalization process until a termination condition is satisfied, wherein performing the one or more iterations comprises:
selecting a weak classifier from a dictionary of weak classifiers,
including the selected weak classifier in the active weak classifier matrix,
performing quantum annealing with the quantum annealing chip to optimize a primal problem based, in part, on the training examples, the active weak classifier matrix, and a plurality of classifier weights, to select one or more weak classifiers with respective discrete weights greater than a threshold, wherein the classifier weights are mapped to the qubits of the quantum annealing chip, and
optimizing 1-regularized risk using the identified one or more weak classifiers and the training to provide a set of continuous classifier weights for the selected active weak classifiers, and
determining whether the termination condition is satisfied, wherein determining whether the termination condition is satisfied comprises determining whether a maximum number of iterations has been performed, wherein when the maximum number of iterations has been performed, the process comprises storing in memory the active weak classifier matrix and the set of continuous classifier weights, and wherein when the maximum number of iterations has not been performed, performing another iteration of the totally corrective boosting with cardinality penalization process.
8. The quantum-computer implemented method of claim 7 , wherein quantum annealing is performed on an Ising model or a restricted Ising model.
9. The quantum-computer implemented method of claim 7 , wherein the plurality of qubits comprise superconducting loops.
10. The quantum-computer implemented method of claim 7 , wherein each unit cell comprises 8 qubits.
11. The quantum-computer implemented method of claim 7 , wherein the programmable couplers comprise inductive couplers.
12. The quantum-computer implemented method of claim 7 , wherein optimizing the primal problem includes using, wherein each of the discrete weights is a non negative fixed point n-bit value, and each bit of each of the n-bit values is mapped to a corresponding quantum bit of the programmable quantum annealing chip.