(In Basis: Lab Development and Application of Data Mining and Learning Systems: Neural ODEs)

Preliminary Meeting Slides (pw: mlai)   Write Ramsés and Patrick ASAP, if you want to participate. Important: Remember to officially register in BASIS as well!

The traditional machine learning paradigm fits individual, specialized models for every dataset anew.  Recent advances in deep learning have enabled an attractive alternative: Amortized Inference.  In this paradigm, large neural networks are pre-trained to process entire datasets to solve a machine learning problem. Thus, laborious retraining of individual models is replaced by a single forward pass of a pre-trained network. The amortized paradigm has already been applied to many areas, including tabular data tasks, causal inference, Bayesian optimization and dynamical system discovery

The goal of this lab is to understand the mechanism of these amortized methods. In controlled experiments, we will investigate the inference algorithms they learn and how they are implemented by the neural network. 

We ask you to present intermediate results during semi-regular meetings and compile your results into a final report. Depending on general interest, you may work on these projects in groups or on your own. 

We will reintroduce the amortized inference paradigm, provide an overview of the experiments we want you to run and more information about the structure of this lab during a preliminary meeting. The preliminary meeting will be held in-person at 2pm c.t. on 15.04.26 in Room 3.110 at the BIT.  

Requirements: working knowledge of machine learning libraries like PyTorch or Jax.