Interventional Style Transfer (IST) for improving model generalization

This technology is a machine learning approach for imputing observations across different observational environments, which can be used to enhance models’ performance on data that falls outside the distribution of the training data.

Unmet Need: Effective method to mitigate confounding effects in model training

Artificial intelligence (AI) and machine learning (ML) models tend to learn spurious correlations between concepts of interest and the context in which relevant examples tend to be observable, i.e. their observational environments. Therefore models perform worse when evaluated on data that contradicts these correlations. Despite the importance of models’ ability to generalize to observations made in different environments and contexts, designing test datasets to evaluate model generalizability has remained difficult, and current existing methods for mitigating confounding by spurious correlations have been ineffective in yielding models that generalize to out-of-distribution (OOD) data.

The Technology: Generation of interventional training distributions for enhanced model generalization

This technology presents an approach, termed Interventional Style Transfer (IST), to diminish the confounding effects of observational environments on feature learning during machine learning model training. It uses a generative model to faithfully impute observations as if they had been collected in different observational environments. In the yielded interventional training distribution, observations thus effectively exist in every observational environment, removing spurious correlations between them and relevant concepts of interest. To evaluate OOD generalization in models, hierarchies within the causal data generation process can be leveraged to establish increasingly challenging OOD generalization tasks. When trained on IST training distributions, computer vision models exhibited major improvements in learning causal representations across these OOD generalization tasks.

This technology has been validated with single-cell fluorescent microscopy datasets.

Applications:

  • Microscopic image classification
  • Analysis of biomedical imaging data, such as microscopy images and x-rays for disease diagnosis
  • Training and evaluation of robust computer vision models
  • Training disentangled conditional generative models
  • Robust autonomous and egocentric vision

Advantages:

  • Generally applicable to existing computer vision models
  • Effective in improving model generalization
  • Yields meaningful, causal representations
  • Compatible with existing imaging data

Lead Inventor:

Wolfgang M. Pernice

Patent Information:

Patent Pending (WO/2024/249579)

Related Publications:

Tech Ventures Reference: