Synthetic Observability Data Generation using GANs
* Register (or log in) to the AI4G Neural Network to add this session to your agenda or watch the replay
Generative modeling is an unsupervised learning technique in machine learning. It involves automatically discovering, and learning, the patterns in the input data. Once learns, it can be used to generate new examples that would be similar to the original dataset. Generative Adversarial Networks, or GANs for short, are an approach to generative modeling using deep learning methods. The challenge (Synthetic Observability Data Generation using GANs) proposes to use GANs for generating infrastructure time-series data (data that has time dependency). Using GANs to generate datasets that should preserve temporal dynamics is challenging compared to generating Images, for example.
This talk will provide a detailed overview of GANs, covering some hands-on exercises using Tensorflow. The talk will also include explanation of GANs for time-series Data, taking TGANs as example. The talk will begin with a quick survey of existing GANs for Time-Series Data, and end with a discussion on possible GANs to consider for this challenge.
Content of the talk:
- GANs and Time-Series Data: State of Art Survey
- Difference between generative and discriminative models
- Identify the problems that GANs can solve
- Understand the roles of the generator and discriminator in a GAN system.
- Understand the advantages and disadvantages of common GAN loss functions.
- Identify possible solutions to common problems with GAN training.
- Use the TensorFlow GAN library to make a GAN – Considering Time-Series Data.
- Discuss the ‘possible’ GANs for this Challenge.