Novel quantum ensemble learning method suitable for multi-classification problem

Abstract

The invention relates to the technical field of quantum machine learning, and discloses a novel quantum ensemble learning method suitable for a multi-classification problem, and the method comprises the steps: initializing the sample weights of all images; using a basis learning algorithm to train the quantum classifier into a weak classifier according to the sample weight vector, and obtaining a weak classifier prediction label and a corresponding prediction probability of the image through projection value measurement of a quantum state; calculating the weight of the weak classifier; updating the sample weight of the image based on the normalization coefficient and the prediction probability of the weak classifier prediction label; and performing weighted summation on the weak classifiers based on the weights of the weak classifiers to obtain strong classifiers, and inputting the images into the strong classifiers to obtain corresponding strong classifier prediction labels. According to the method, classifier weights and sample weights are separated, adjustment factors are added, the degree of freedom of the whole algorithm is improved, the accuracy requirement for weak classifiers is lowered, the adaptation range is wider, and it is ensured that the accuracy of the total classifier is improved along with increase of the number of the classifiers.

Claims

1. A novel quantum ensemble learning method suitable for multi-classification problems, characterized by: It includes the following steps: S1. Initialize the sample weights of all images to form a sample weight vector, where all sample weights in the sample weight vector have the same initial value; set the iteration number as t, and initially t = 1; S2. Use the base learning algorithm to train the quantum classifier into the t-th weak classifier according to the sample weight vector, and obtain the prediction label of the weak classifier for the image and the corresponding prediction probability through the projection value measurement of the quantum state; S3. Calculate the weight of the t-th weak classifier based on the adjustment factor and the sum of the weighted confidences of the t-th weak classifier on the correctly classified and misclassified image samples; S4. Update the sample weights of the image based on the normalization coefficient and the prediction probability of the prediction label of the weak classifier; S5. If t < T, let t = t + 1, and repeat steps S2 to S4; T is the total number of weak classifiers; S6. Perform a weighted sum of the T weak classifiers based on the weights of the weak classifiers to obtain a strong classifier, input the image into the strong classifier, and obtain the corresponding prediction label of the strong classifier.

2. A novel quantum ensemble learning method suitable for multi-classification problems according to claim 1, characterized in that: The initialization of the sample weights of all images to form a sample weight vector, where all sample weights in the sample weight vector have the same initial value, specifically includes: The sample weight vector corresponding to the t-th weak classifier w t,i is the sample weight of the i-th image used to train the t-th weak classifier, w 1,i =1 / N; N is the total number of images.

3. A novel quantum ensemble learning method suitable for multi-classification problems according to claim 1, characterized in that: The use of the base learning algorithm to train the quantum classifier into the t-th weak classifier according to the sample weight vector, and obtain the prediction label of the weak classifier for the image and the corresponding prediction probability through the projection value measurement of the quantum state, specifically includes: Determine the number m of qubits for image classification, input the image into T weak classifiers based on the quantum classifier respectively, and each weak classifier outputs the measurement results of m qubits; Pass 2 m The base projector measures the projection value of the m qubits under the calculation basis to obtain the corresponding measurement state; m The base projectors are evenly divided into K groups, where K is the total number of image categories; If the measurement state is within the index range of the kth group of base projectors, the image corresponding to the measurement state is classified into the kth class; the predicted probability that the i-th image passes through the t-th weak classifier and is classified into the k-th class is P k (x i ,θ t ), x i represents the i-th image, θ t is the parameter of the tth weak classifier.

4. A novel quantum ensemble learning method suitable for multi-classification problems according to claim 3, characterized in that: The determination of the number of qubits for image classification, specifically includes: where m is the number of qubits.

5. A novel quantum ensemble learning method suitable for multi-classification problems according to claim 3, characterized in that: The 2 m The base projectors are divided into K groups, including: The index range of the k-th group of base projectors arrive Discard the last A base projector.

6. A novel quantum ensemble learning method suitable for multi-classification problems according to claim 3, characterized in that: The predicted probability that the i-th image passes through the t-th weak classifier and is classified as the k-th class is P k (x i ,θ t ), specifically including: Among them, x i represents the i-th image, θ t is the parameter of the t-th weak classifier, Tr(·) represents the trace operation, ρ(x i ,θ t ) represents the parameter θ t The reduced density matrix of m qubits measured by the parameterized quantum classifier; Π j represents the j-th basis projector.

7. A novel quantum ensemble learning method suitable for multi-classification problems according to claim 1, characterized in that: The calculation of the weight of the t-th weak classifier based on the adjustment factor and the sum of the weighted confidences of the t-th weak classifier on the correctly classified and misclassified image samples, specifically includes: Among them, α t is the weight of the tth weak classifier; c t is the adjustment factor of the t-th weak classifier: The strong classifier prediction label output by the strong classifier composed of all the weak classifiers trained so far, y i is the true classification label of the i-th image; δ is the Dirac symbol, c represents a real variable, represents traversing the real number c so that c t Taking c can maximize the classification accuracy; represents the sum of the weighted confidence of the t-th weak classifier on the correctly classified image samples, It represents the sum of weighted confidences of the t-th weak classifier on the misclassified image samples.

8. A novel quantum ensemble learning method suitable for multi-classification problems according to claim 7, characterized in that: described represents the sum of the weighted confidence of the t-th weak classifier on the correctly classified image samples, Represents the sum of the weighted confidence of the t-th weak classifier on the misclassified image samples, including: y i is the true classification label of the i-th image, is the weak classifier prediction label output by the t-th weak classifier for the i-th image, δ is the Dirac symbol, for The prediction confidence of Represents the final parameters of the t-th weak classifier after training.

9. A novel quantum ensemble learning method suitable for multi-classification problems according to claim 1, characterized in that: The update of the sample weights of the image based on the normalization coefficient and the prediction probability of the prediction label of the weak classifier, specifically includes: Among them, the normalization coefficient w t is the sample weight vector corresponding to the tth weak classifier, y i is the true classification label of the i-th image, is the weak classifier prediction label output by the t-th weak classifier for the i-th image, δ is the Dirac symbol, for The prediction confidence of Represents the final parameters of the t-th weak classifier after training.

10. A novel quantum ensemble learning method applicable to multi-classification problems according to claim 1, characterized in that: The performance of a weighted sum of the T weak classifiers based on the weights of the weak classifiers to obtain a strong classifier, input the image into the strong classifier, and obtain the corresponding prediction label of the strong classifier, specifically includes: represents a strong classifier, α t is the weight of the t-th weak classifier, represents the tth weak classifier, x represents the image, Represents the final parameters of the t-th weak classifier after training; Represents the strong classifier prediction label of the i-th image.

Images