Reproduce “Discovering interpretable models of scientific image data with deep learning” on OpenResearch

About this paper

Christopher J. Soelistyo, Alan R. Lowe

Researchers developed a framework combining disentangled representation learning, sparse neural networks, and symbolic regression to create interpretable deep learning models for scientific image analysis. Applying this to cell phase classification from microscopy images, their models achieved 97-99% accuracy while providing explicit, human-understandable insights into the biological features driving classification, such as cell size and eccentricity.

Key insights

• Interpretability in deep learning for scientific data is achievable without significant performance compromise (only 2-3% accuracy difference from black-box models).
• Disentangled latent spaces can learn and isolate biologically meaningful features, such as cell size and eccentricity, directly relevant to classification tasks.
• Models designed for interpretability also exhibit greater domain-appropriateness and robustness to adversarial attacks by focusing on relevant morphological features.

Summary by alphaXiv.