AI and machine learning
Machine learning that is tested before it is trusted. Tabular models already work inside SCALAR; research continues into neural networks and computer vision for scientific images.
Available in SCALAR
- A neural network and four other models for tabular data
- k-fold cross-validation, so every sample is predicted by a model that never saw it
- A shuffled-label baseline that shows what chance alone would give
- Permutation importance and a local library of saved models
Ongoing research
- Neural network architectures and deep learning
- Computer vision and image segmentation
- Fluorescence microscopy, cell and tissue image analysis
Wider research areas
- Model optimisation and weight-space programming
- Predictive modelling
- AI-assisted scientific workflows
Research results will be published here when they have been validated. Nothing on this page has been clinically validated.
Train a network in your browser
A small neural network learns to separate two classes of points. Watch the boundary form, and compare the loss on the training points with the loss on points it never sees.
Interactive demo with synthetic data: a fully connected network with tanh units, trained with Adam in your browser. It illustrates training and validation; it is not one of SIGMA IV's research models.
Find the cells in an image
Segmentation turns an image into objects that can be counted and measured. Here with the classical method: smooth, choose a threshold, and group the bright pixels that touch.
Interactive demo with a synthetic image. It uses a global threshold (Otsu's method or your own) and connected components, not a trained network. Where two cells touch they merge into one object: one of the problems learned segmentation models are studied for.
Machine learning in SCALAR
Researchers can train models on their own table, one row per sample, and see whether a model finds anything beyond chance. Everything runs and stays on their own computer.
With fewer than about a hundred samples, a random forest or linear model is usually as good as a neural network, and SCALAR says so.
More about SCALAR