We aim to find a good way to automatically find architectures of neural networks that represent underlying connections in the Universe of the data. In Biology, for example, these are usually termed "Biologically-Informed neural networks" or sometimes "Visual Neural Networks". They are built to contain Gene-Pathway or Protein-Interaction information.
More generally, we consider discovery any hierarchical structure for which observed features (genes) are the leaf nodes. This is different from typical graph neural network problems, because the branch nodes (the pathways and processes) are not directly observed, but form hidden nodes or layers (including with skip connections) in the neural network.
We had a Poster for this purpose on the JGU Workshop for AI for Scientific Discovery: Interpretable AutoML for biological knowledge discovery from multi-modal data. The code here was used for the results shown on this poster. Further work will, be simplifying the available code and streamlining the input of hierarchical structure to build NNs. Likely a functional instead of an object oriented approach would be more accessible.
- Build minimal running example of cancer-net (pytorch implementation of P-NET)
- Get P-NET Dataset
- Search for AutoML options, that provide single edge optimization
- Simplify hierachical structure.
Experiments with different architectures:
- PNET, but graph is randomized
- Fully connected net
- both with and without PNets
- introduce spareseness by regularization (done but not finished)
- Introduce spareseness by mask: [ ] random [ ] with hierarcical network.