Autoassociative Learning of Structural Representations for Modeling and Classification in Medical Imaging

Deep learning architectures based on convolutional neural networks tend to rely on continuous, smooth features. While this characteristics provides significant robustness and proves useful in many real-world tasks, it is strikingly incompatible with the physical characteristic of the world, which, at the scale in which humans operate, comprises crisp objects, typically representing well-defined categories. This study proposes a class of neurosymbolic systems that learn by reconstructing images in terms of visual primitives and are thus forced to form high-level, structural explanations of them. When applied to the task of diagnosing abnormalities in histological imaging, the method proved superior to a conventional deep learning architecture in terms of classification accuracy, while being more transparent.

💡 Research Summary

The paper introduces a novel neurosymbolic autoencoder called ASR (Auto‑associative Structural Representations) that learns to reconstruct images using a set of interpretable visual primitives rather than raw pixel intensities. The authors argue that conventional convolutional neural networks (CNNs) excel at extracting smooth, continuous features but fail to capture the discrete, object‑centric nature of the visual world, leading to over‑parameterisation, data inefficiency, and poor explainability—issues especially acute in medical imaging where annotated data are scarce and interpretability is crucial.

ASR’s architecture consists of three main components: (1) a multi‑scale convolutional encoder built from a stack of ConvBlocks that extracts hierarchical feature maps and predicts a uniform background colour; (2) a series of Modelers, one per spatial scale, that map each latent feature vector to six parameters defining an ellipse (horizontal scale w, vertical scale h, rotation d, and RGB colour a). The Modelers are implemented as 1×1 convolutions with sigmoid activations and linear rescaling to enforce meaningful bounds (e.g., w, h ∈