Neural Production Systems (NPS) Implementation and Augmentation

This repository contains an implementation and augmentation of Neural Production Systems (NPS) as introduced in NPS paper (https://arxiv.org/pdf/2103.01937.pdf). The code is built upon the NPS paper repository (GitHub), as well as CausalMBRL repository (https://github.com/dido1998/CausalMBRL.git) for the physics experiment. For a detailed analysis of the experiments, refer to report.pdf. The repository also contains trained models and output images.

Contents

Arithmetic

MNIST

In the MNIST folder, execute the following commands:

pip install -r requirements.txt
pip install -e git+https://github.com/ncullen93/torchsample.git@ea4d1b3975f68be0521941e733887ed667a1b46e#egg=torchsample

Then, run inside each experiment directory:

./run.sh seed

Six Transformations (six_transMNIST):

• Evaluate NPS for 6 MNIST transformations
• Introduce exploration phase

Composite Transformations (compMNIST, compMNIST_step):

• Two different implementations of sequential NPS
• Take one image-digit and apply a composite transformation
• NPS manages to partially solve the task without overfitting in compMNIST_step and fully solve it by overfitting in compMNIST.

Double Digit MNIST (catMNIST, catMNISTv2, catMNISTvseq):

• Take two or three digit-images as input and output a single, double, or triple digit
• NPS performs well on the first two settings but fails on the sequential one (catMNISTvseq)

Customizing Transformations (cust_transMNIST):

• Dataset contains image transformation of specific degrees and pixels
• NPS manages to create 4 adjustable rules that can rotate or translate digit-images at different scales

Physics

We integrate the NPS algorithm into the physics experiment as introduced in (https://arxiv.org/abs/2107.00848).

References

[1] NPS Paper
[2] NPS Paper Repository
[3] Physics Experiment Repository
[4] Physics Experiment Paper