TWCC22_PPT_v3_KL.pdf quantum computer, quantum computer , quantum computer , quantum computer,

TWCC22_PPT_v3_KL.pdf quantum computer, quantum computer , quantum computer , quantum computer,

• 1.
  Classification of tumormetastasis data using Quantum kernel-based algorithm Team：QKBA Mentors：Leo (NVIDIA), Jay (NVIDIA) Mentees：Tai-Yue Li (National Synchrotron Radiation Research Center), Cheng-Fang Su (National Yang Ming Chiao Tung University),, Chung-Hsing Hsu (Asia University), Kuan-Cheng Chen (QuEST), Ka-Lok Ng (Asia University) tim312508@gmail.com , scf1204@nycu.edu.tw
• 2.
  Introduction Primary tumor Bloodor lymphatic system New tumor High mortality Cancer cells We need an effective method to predict tumor metastasis to help patients prevent it in advance.
• 3.
  Introduction non-metastasis tumor metastasis tumor NextGeneration Sequencing (NGS) … non-metastasis tumor gene … metastasis tumor gene Biomarker
• 4.
  Introduction Currently, we don'thave effective methods to find biomarkers of tumor metastasis. How effective is the classification of tumor metastasis data using the QSVM algorithm? Supervised learning with quantum enhanced feature spaces Nature volume 567, pages209–212 (2019)
• 5.
  Introduction x1 |0 > |0 > classicaldata space 2 data points with 2 features： x2 x1 feature1 x1 feature2 x2 feature1 x2 feature2 x2 x1 x1 feature1 x1 feature2 U(x1) x2 feature1 x2 feature2 U+(x2) q1 q2 Quantum feature space x2 x1 n features = n qubits Quantum feature map circuit (non-liner map function) Transition amplitude of two data a x1 x2 Measure the expectations value at zero state Computation time costs = 2n (n qubits) from numbers of features 0.5m2 (m data) from numbers of data x 1 a 1 a x1 x2 x1 x2 SVM Classification Quantum kernel matrix
• 6.
  Issue Insufficient computing resourcesand long computing time are the main problems. We screened out 48 important gene expression as features for tumor metastasis classification. Ø The amount of calculation required is as high as 2 to the 48th power. (248) Classical computer • Difficult to perform Real quantum computer • Calculation cost very high • Result will be affected by noise Although the principal component analysis (PCA) can reduce the number of qubits can make the program executable on a classical computer. Ø The number of qubits and the amount of data increase, the calculation time will also increase significantly. Ø We need to know which gene is related to tumor metastasis. Therefore, a better solution is to alleviate the limitation of computing resources
• 7.
  Solution We use thecuQuantum SDK to accelerate the GPU simulate QSVM. And cuQuantum can use the two technologies of cuStateVec and cuTensorNet to accelerate the matrix calculation using GPU. The quantum circuit operations can be regarded as a series of matrix operations. |0 > |0 > ….. U(x1) U+(x2) M |0 > |0 > ….. QSVM quantum circuit ⋯ ⋮ ⋱ ⋮ ⋯ ⋯ ⋮ ⋱ ⋮ ⋯ Series of matrix operations … In this hackathon： 1. Solved the slow calculation problem from Qiskit. 2. Install cuStateVec support GPU quantum computer simulator 3. Using cuTensorNet to solve insufficient memory problem https://developer.nvidia.com/cuquantum-sdk
• 8.
  Result 17 18 1920 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 0.01 0.1 1 10 100 Time (min) No. Qubits 334data-cpu 334data-cuStatevec-A6000 334data-cuStatevec-A100*2 2data-cpu 2data-cuStatevec-A6000 2data-cuStatevec-A100*2 34.1 min 21.7 min QSVM runtime test 3 min 35.8 min 1.9 min 1.1 min 19x speedup 33x speedup 1.6x speedup 11.4x speedup
• 9.
  0 4 812 16 20 24 28 32 36 40 44 48 84 86 88 90 92 94 96 98 100 102 Accuracy (%) Dimention SVM-train QSVM-train QSVM(cost)-train Result Tumor metastasis classification 100% at 8q 100% at 20q 99% at 37q
• 10.
  Future work Ø Wewill complete the code of cuTensorNet running the QSVM quantum circuit. Ø We need to solve the overfitting problem in our study Ø Understand the relationship and limitation of data size, circuit depth and qubit number on QSVM operation time. Data size Circuit depth Qubit number QSVM cuQuantum 0 4 8 12 16 20 24 84 86 88 90 92 94 96 98 100 102 Accuracy Dimention Trian Test
• 11.
  Thanks In this hackathon,we learned much about cuQuantum and GPU, hoping these tools can accelerate quantum computing research in future. Using the cuQuantum SDK to accelerate GPU-simulated quantum computing brings hope to our research.
• 12.
  Result
• 13.
  Result Generative modeling isnot typically used for binary classification tasks, as it is designed to model the probability distribution of the data rather than to classify it directly. However, it is possible to use generative models for binary classification tasks by combining them with a discriminative model. One approach is to use a generative model, such as a Gaussian mixture model or a variational autoencoder, to model the distribution of the data for each class. Then, a discriminative model, such as a logistic regression or a neural network, can be trained on the latent variables generated by the generative model to perform binary classification. Another approach is to use a generative adversarial network (GAN) to generate synthetic data for each class. The discriminator of the GAN can then be trained to classify the real and synthetic data as belonging to one of the two classes. In both cases, it is important to ensure that the generative model is able to capture the underlying distribution of the data for each class accurately. Otherwise, the performance of the discriminative model will be limited.
• 14.
  Result Here are thesteps for the 1st approach on using a generative model and a discriminative model for binary classification: 1.Define the problem: Let's say we want to classify images as either cats or dogs. We have a relatively small dataset of labeled images, and we want to use a generative model to augment the dataset and improve the performance of our classifier. 2.Train the generative model: We will train a generative model, such as a Gaussian mixture model or a variational autoencoder, on the labeled images in our dataset. The generative model will learn the underlying probability distribution of the data for each class (i.e., the distribution of cat images and the distribution of dog images). Once the generative model is trained, it can generate synthetic images of cats and dogs that are similar to the real images in the dataset. 3.Generate synthetic data: Using the trained generative model, we can generate synthetic images of cats and dogs to augment our dataset. We can control the number and type of synthetic images we generate by adjusting the input to the generative model. 4.Train the discriminative model: We will train a discriminative model, such as a logistic regression or a neural network, on the latent variables generated by the generative model. The discriminative model will learn to classify the images as either cats or dogs based on the latent variables. 5.Evaluate the performance: We can evaluate the performance of our classifier on a separate test set to see if the synthetic data has improved its performance. If the performance has improved, we can continue to generate synthetic data and use it to augment our dataset further. Overall, the use of a generative model to model the distribution of the data and generate synthetic data can be a powerful tool for improving the performance of binary classifiers, especially when the original dataset is small or unbalanced.
• 15.
  Result Here's an exampleof the 2nd approach on how to use a GAN for binary classification: 1.Define the problem: Let's say we want to classify images as either cats or dogs. We have a dataset of images of cats and dogs, but it is relatively small, and we want to generate synthetic images to augment the dataset and improve the performance of our classifier. 2.Build the GAN: We will build a GAN consisting of a generator and a discriminator. The generator takes in a random noise vector and outputs a synthetic image, while the discriminator takes in an image and predicts whether it is real (i.e., from the dataset) or fake (i.e., generated by the generator). We will train the generator and discriminator in an adversarial manner, where the generator tries to fool the discriminator, and the discriminator tries to correctly classify the images as real or fake. 3.Train the GAN: We will train the GAN on the images in our dataset until the generator can produce realistic- looking synthetic images of cats and dogs. 4.Generate synthetic data: Once the GAN is trained, we can use the generator to generate synthetic images of cats and dogs. We can control the number and type of synthetic images we generate by adjusting the random noise vector input to the generator. 5.Train the classifier: We can then combine the real and synthetic images to create an augmented dataset and train our binary classifier on this dataset. The classifier should be able to classify both the real and synthetic images accurately. 6.Evaluate the performance: We can evaluate the performance of our classifier on a separate test set to see if the synthetic data has improved its performance. If the performance has improved, we can continue to generate synthetic data and use it to augment our dataset further. Overall, the use of a GAN to generate synthetic data can be a powerful tool for improving the performance of binary classifiers, especially when the original dataset is small or unbalanced.
• 16.
  Result
• 17.
  Result