Motivation

As the name suggests we were motivated by the ever so benevolent Jarvis as shown in one of the most famous marvel movies, Iron-man. Our idea is similar to that of Google assistant, Amazon's Alexa and Apple's Siri. There is generally a very big trade-off between engineering effort and deep learning, where one can achieve high performance with very smart engineering at the cost of generalization to other similar domains. Since all these companies have huge teams for all these products often the high performance rate is achieved because of the smart engineering effort of the individual teams that tackle a very small component of the entire product.

In our project we are trying to design an assistant, which tries to focus more on the deep learning aspect of assistant in the hopes of making a generic well performing AI that can easily expand to other tasks. For this idea we found a paper which talks about different aspects needed to implement a task-oriented dialog system, which can be thought of as an Assistant without the additional feature of user personalization. We believe having a good task-oriented dialog system that can expand to different tasks easily (by hopefully leveraging pre-existing task knowledge), is a good starting point and a stepping stone to the more benevolent Jarvis.

Minimal Viable Action Plan

Project Objectives

• Understand the data representation and pre-process this data to meet our needs.
• Understand the deeper implementation details of each model as discussed in the paper.
• Implement the models using PyTorch
  • Evaluate the individual models (belief tracker, intent network, generation network and policy network) to identify domain sensitivity.
  • Try to improve the architecture of each model to make it domain agnostic.
• Evaluate the overall model using peer evaluation and BLUE scoring technique.
• Compare how the model performs against other sophisticated architectures with similar goals.(NBT)
• Compare task wise performance of the model on independent domain specific tasks

Methodologies

• Implement the base line model as described in paper
  • The current implementation is written in Theano.
    • It is difficult to expand on it since we would need to implemented all operators from scratch.
    • Contains a lot of unnecessary code that can be avoided by leveraging existing frameworks like PyTorch.
  • Implement the basic version using Pytorch.
• Extend the implementation to other domains independently.
  • Use the existing datasets and evaluate our model on it.

Stretch Goals:

• Combine different domains to train a single model for multiple tasks.