top of page
Image by DeepMind

Core Technology

Hylomorph's Structural Machine Learning is a state-of-the-art neurosymbolic algorithm: a combination of formal and heuristic approaches to machine learning and program synthesis.
  • Can incorporate domain-specific knowledge
  • Produces human-interpretable results
  • Can learn from small data
  • Gives much better generalization

SML improves on Deep Learning


Our Structural Machine Learning has successful industrial application in:
  • Time-series analysis
  • Biomedicine
  • Digital Signal Processing
  • Pulmonary Disease Detection
Results are superior to deep learning in all key metrics: training time, generalization, explainability.

Structural Machine Learning models can be compiled to run efficiently on target hardware (e.g. microcontrollers/edge computing devices).

Case study: Molecular discovery

Structured Machine Learning excells at molecular discovery, delivering high-quality predictions of molecules in equilibrium. Our method improves on the state-of-the-art (Zhang et al. 2020), significantly reducing the error of estimation of zero-point vibrational energy (ZPVE) and electronic spatial extent (<R^2>) on the QM9 benchmark (Ruddigkeit et al. 2012, Ramakrishnan et al. 2014). We achieve this using a hybrid neuro-symbolic model, which augments a Graph Neural Network (GNN) with domain-aware features.
L. Ruddigkeit, R. van Deursen, L. C. Blum, J.-L. Reymond, Enumeration of 166 billion organic small molecules in the chemical universe database GDB-17, J. Chem. Inf. Model. 52, 2864–2875, 2012.
R. Ramakrishnan, P. O. Dral, M. Rupp, O. A. von Lilienfeld, Quantum chemistry structures and properties of 134 kilo molecules, Scientific Data 1, 140022, 2014.
S. Zhang, Y. Liu, and L. Xie, “Molecular Mechanics-Driven Graph Neural Network with Multiplex Graph for Molecular Structures.” arXiv, Nov. 15, 2020. doi: 10.48550/arXiv.2011.07457.​

Case study: Detection of a pulmonary disease

In collaboration with a client from the USA, we applied a variant of our Structural Machine Learning (SML) to the task of detecting a pulmonary disease based on time series collected by a wideband e-stethoscope. SML synthesised a model in a bespoke Domain-Specific Language using a very small training set (tens of cases). The synthesised model generalized well beyond the training set. The explainable SML model is sufficiently informative that it affords insight to clinicians and device manufacturers. 
bottom of page