Inferential Learning
Table of Contents
- System's knowledge bases is represented in a set of beliefs
- Learning is updating the topological structures as well as weights of those beliefs
- Updates is done through inference rules
- The system can be queried for answers.
For contrast with Algorithmic Learning, see Different Conceptions of Learning.
In inferential learning there is no distinction between reasoning and learning. Only thing is: Reasoning is temporary while learning is stored.
1. Thoughts
Some thoughts on the limitations (and their possible solutions) of Inferential Learning
1.1. High Dimensional Input
- How to represent high dimensional input
- How to represent action statements:
- Ram is Human: Ram -> Human (S -> P(t)) [OK]
Ram ate rice: Ram ?? Rice [Not OK]
There are infinite actions. Even if we represent them how to we write inference rules for them?
- Represent action verbs as vectors, then Inference rules as ??
- Represent statements as vectors, then inference rules as functions
- Can we use vectors (or functions too??) as statements? Because statements are discrete but vectors can represent infinites
Then inference rule can be represented by a function f:
f(S) -> {(S', t)}
Given an Statement S, f gives a new statements (S' is statement, t is its belief)
f(S, z) -> (S', t)
1.2. Temporal relations among statements
As time goes on system observes many things. Those observation are input to the system as statements. How can temporal relation among such statements be established? How can causation be established?