Neural Production System

Unofficial implementation of paper NeurIPS 2021 paper "Neural Production Systems" (NPS).

Features

• supports three experiments provided by official source code:
  • coordinate_arithmetic
  • sequential_arithmetic
  • mnist_transformation
• thoroughly refactored and re-implemented these models and datasets
  • remove more than 90% of code
  • refactored the building of models and datasets to the same style
  • boosted training and testing speed to some degree
• (hope) corrected almost all the gaps between the paper declaration and its official source code:
  • official q/k attention dose not use shared weights (partially corrected)
  • official "hidden states" dose not obtained by sequential model (todo)
  • parimay and contextual selections are not conducted as its algorithm describes (partially)

References

• published paper and official code 1: NeurIPS 2021 Supplemental
• arXiv paper and official code 2: anirudh9119 / neural_production_systems

Comments

Do not hesitate to contact me if you want to discuss some ideas related to NPS with someone.