T-720-ATAI-2022 Main

ATAI-22 Reykjavik University



Final Project:


Aim: The final project aims at giving you a deeper insight into an AGI-aspiring system (i.e. ONA) and its analogy making capabilities. This includes inference of similarities through interaction, as well as negative knowledge transfer.

Summary: Create your own example with all three types of knowledge transfer where ONA learns about the information necessary to make analogies. Analyze the deduction, abduction, and induction abilities of ONA when using analogies and show the influence of negative knowledge transfer on the confidence and performance of AERA.

Create your own example

For the final project, you need to develop your own example showing all three knowledge transfer types: 1) Knowledge transfer via shared property, 2) knowledge transfer via identical relation, and 3) knowledge transfer via comparison. This time, you have to make ONA learn about different phenomena using environmental feedback. For this, ONA needs seed information (i.e., a small amount of base knowledge that can be used to 1) interact with the environment and 2) make analogies with).

A (very) short example to give you an idea:

Object A is hard and orange. If you pick Object A, it bounces. Something that bounces can be used as a ball to play basketball with.
Object B is hard and blue. If you pick object B, it bounces as well.
You see object C, which is hard and green. The goal is to play basketball.

Objects C, D, and E: C is harder than D, D is harder than E. Object D bounces.
Can object C be used as a basketball? Can object E be used as a basketball?
Or, given objects C, D, and E, set the goal to play basketball. Which ball does ONA pick?

Analyze ONA's reasoning

With this example, show ONA's deduction, abduction, and induction abilities by asking questions about the example.

  1. Ask similar questions to those from engineering assignment 2 tuned to your example to get an insight into the deduction and induction capabilities.
  2. Especially put a focus on the abduction abilities of ONA. On GitHub, you can find the duck.nal and the circus.nal examples to give you an idea how to approach this. Use similar questions about your example that show how ONA can abduce properties and relations.

Additionally, find out what the minimal amount of information given to ONA initially, that leads to successfully answering your questions, is.

Negative knowledge transfer

Negative knowledge transfer happens if solving an earlier task hinders solving a new one. This occurs, for example, if the system has learned a particular relation in the first task, which does not hold in the second, even though both provide a similar observation. Think, for example, of driving on the other side of the road. Even though most observations are similar (or even the same) as before, some relations are different (e.g., the indicators on the right of the steering wheel instead of the left). Negative knowledge transfer might lead to the misuse of the indicators when turning.

In the previous example, this might be another hard object that is a cube that does not bounce when picked up.

How does negative transfer of knowledge for a particular relation influence the confidence in this relation?

Report

Write a report about your findings. Describe how your example shows the different types of knowledge transfer. Include the deduction, induction, and abduction abilities. Also include your .nal files when uploading your solution.