Project

Generative adversarial network for palliative care

Project (M.S., Computer Science)--California State University, Sacramento, 2017.

In this body of work, we look at implementing the Generative Adversarial Network to a palliative care dataset. The implementation requires that our model, which we call the Palliative Care Generative Adversarial Network(PCGAN), train on a relatively small dataset containing outliers from missing information. This problem is not entirely unique to our model. Many real-world applications simply do not have large or complete datasets for training. To address these issues, the PCGAN implements Convolutional Neural Networks to get a rich representation from a 1-dimensional continuous real-world dataset. We explore the use of latent space parameterization to overcome the dataset size problem. In addition, PCGAN implements techniques such as feature matching, batch normalization, one-sided label smoothing to circumvent stability concerns. We train our dataset with minimal data preprocessing and avoid the need for a conversion matrix. The trained PCGAN model shows relatively good results in predicting a 6-month prognosis and produces realistic samples. PCGAN shows that such an approach is indeed a viable solution to the dataset size problem, and learning feature representations resilient to outliers.

In this body of work, we look at implementing the Generative Adversarial Network to a palliative care dataset. The implementation requires that our model, which we call the Palliative Care Generative Adversarial Network(PCGAN), train on a relatively small dataset containing outliers from missing information. This problem is not entirely unique to our model. Many real-world applications simply do not have large or complete datasets for training. To address these issues, the PCGAN implements Convolutional Neural Networks to get a rich representation from a 1-dimensional continuous real-world dataset. We explore the use of latent space parameterization to overcome the dataset size problem. In addition, PCGAN implements techniques such as feature matching, batch normalization, one-sided label smoothing to circumvent stability concerns. We train our dataset with minimal data preprocessing and avoid the need for a conversion matrix. The trained PCGAN model shows relatively good results in predicting a 6-month prognosis and produces realistic samples. PCGAN shows that such an approach is indeed a viable solution to the dataset size problem, and learning feature representations resilient to outliers.

Relationships

Items