Prior for Value-Aligned Agent

National Action Council for Minorities in Engineering(NACME) Google Applied Machine Learning Intensive (AMLI) at the University of Kentucky

Team

• Md Sultan Al Nahian - University of Kentucky (Research Mentor)
• Elizabeth Eyeson - University of Colorado Boulder
• Kavisca Kuruparanantha - University of Kentucky
• Steven Aponte - Stevens Institute of Technology

Description

Goal-driven AI is susceptible to neglecting ethical concerns due to its blind prioritization of optimization in order to accomplish its goal with maximal performance. This project is a prior to a value-aligned agent that will be taught human values such that it will be trained to take actions that closer align with "normative" human behavior.

Model Tasks

• Classification: Actions are classified as either normative or non-normative. Such classification is indicated by either a 0 (non-normative) or 1 (normative).
• Sequence Generation: Sequences describing the intentions behind actions taken are generated.

Model Architectures

Architecture With Bidirectional LSTM

• Implemented in bidirectional-lstm.ipynb
• Performs classification task

DPCNN

• Implemented in DPCNN-master
• Developed by Rje Johnson and Tong Zhong (original repository)
• Performs classification task

Seq2Seq

• Implemented in seq-seq-learning-tensorflow.ipynb
• Derived from Tensorflow documentation (original documentation)
• Performs sequence generation task

Requirements

• Command to install required libraries: pip install -r /path/to/requirements.txt
• DPCNN code requires TensorFlow 1.14-1.19 and Python 3.6
• Bidirectional LSTM and Seq2Seq code require either Jupyter or Colab
• Seq2Seq code requires TensorFlow 2.0

Usage Instructions

Architecture With Bidirectional LSTM

• Run bidirectional-lstm.ipynb in either Jupyter or Colab

DPCNN

• Run run.py in ./DPCNN-master/DPCNN-master in IDE supporting Python

Seq2Seq

• Run seq-seq-learning-tensorflow.ipynb in either Jupyter or Colab