Informed Meta-Learning with INPs

This repository contains the code to reproduce the results of the exepriments presented in the paper:

Towards Automated Knowledge Integration From Human-Interpretable Representations published at ICLR 2025

For citations, use the following:

@inproceedings{
kobalczyk2025towards,
title={Towards Automated Knowledge Integration From Human-Interpretable Representations},
author={Katarzyna Kobalczyk and Mihaela van der Schaar},
booktitle={The Thirteenth International Conference on Learning Representations},
year={2025},
url={https://openreview.net/forum?id=NTHMw8S1Ow}
}

Setup

The environment.yaml lists the required packages to reproduce the experiments presented in the paper. To install the evnironment run:

conda env create -f environment.yaml

conda activate inps

Experiments with Synthetic Data

jobs/run_sinusoids.sh contais commands that need to be run to reproduce the experiments with synthetic data

After training the models, results can be analyzed with the following two notebooks:

• evaluation/evaluate_sinusoids.ipynb contains the analysis of the base experiments
• evaluation/evaluate_sinusoids_dist_shift.ipynb contains the analysis of the train/test distribution shift experiment

Experiments with the Tempereatures Dataset

jobs/run_temperatures.sh contais commands that need to be run to reproduce the experiments with the tempereatures datasets

After training the models, results can be analyzes with evaluation/evaluate_temperature.ipynb