Probabilistic Numeric CNNs (PNCNNs)

This repository contains the implementation and experiments for the paper presented at ICLR2021.

Marc Finzi¹, Roberto Bondesan², Max Welling², "Probabilistic Numeric Convolutional Neural Networks", ICLR 2021. [arxiv]

¹ New York University (Work done during internship at Qualcomm AI Research) ² Qualcomm AI Research (Qualcomm AI Research is an initiative of Qualcomm Technologies, Inc.)

Training the Models

Assuming the docker container has been built successfully, you can train the relevant models with the following commands. The author was able to execute his code in a Docker container based on the image nvidia/cuda:11.0.3-base-ubuntu20.04. After installing using setup.py, you can train the relevant models with the following commands.

SuperPixelMNIST-75

Train the PNCNN on SuperPixelMNIST-75, a variant of MNIST digit classification where images are replaced by a collection of 75 irregularly spaced superpixels. The dataset will be automatically downloaded when running the script for the first time to the directory specified by --data_dir which defaults to scripts/datasets.

python scripts/train_superpixel.py

The script contains several command line arguments such as --data_dir --bs --lr --num_epochs that can be specified. To see all command line arguments and their default values, run python scripts/train_superpixel.py --help. As shown in the table below, PNCNN greatly improves over the previous state of the art performance.

Reference

If you find our work useful, please cite

@inproceedings{finzi2021,  
  title={Probabilistic Numeric Convolutional Neural Networks},  
  author={Marc Anton Finzi, Roberto Bondesan, Max Welling},  
  booktitle={International Conference on Learning Representations},  
  year={2021},  
  url={https://openreview.net/forum?id=T1XmO8ScKim}  
}