merge-datasets-gan

A Generative Adversarial Network (GAN) for learning a shared representation for different datasets

Background

• Alzheimer’s Disease (AD) is one of the most widespread and costliest illnesses in first world countries
• No treatment available to permanently cure AD
  • However, effects of AD progresses in stages
  • Gravity of symptoms increases incrementally as time passes
• It’s crucial to identify AD as soon as possible to mitigate symptoms and test new treatments (which mostly apply to the first stages only)
• Screening tools based on Machine Learning (ML) show promising accuracy and increased efficiency in terms of time and human-resources

Motivation

• Multiple different datasets for AD classification exist (e.g., CANARY, DementiaBank, etc.)
  • Them all have differences (e.g., data collection strategy, modality, etc.)
• All are relatively small for training ML and especially DL models effectively
• There is a lack of large scale datasets related to AD hindering progress of research on automatic AD screening

Proposed Approach and Method

• Try to combine data stemming from different sources (datasets) building a shared representation that can be fed to a classifier
  • Very ambitious goal (high risk but high reward)
• How do we build the shared representation? Using a GAN!
  • Train encoder to learn representations of data from different datasets
  • Train discriminator to recognize from which dataset the representation was originally derived
  • When the discriminator cannot make proper distinction anymore then it means representations are sufficiently similar (i.e., homogeneous) and can be used all together as data points for classifier
    • Note: do not want all our data points to be equal, crucial to learn homogeneous representation preserving individual info from each datapoint (big challenge!)

GAN Architecture for Shared Representation Learning