Elephant is an open-source project on how to train machine understanding. The method features:
- Step-by-step training from simple to complex
- Making use of automated hypotheses generation and trial
- Adapting to user-defined languages of subject domains
- The continued evolution of knowledge for solving problems
What Is Understanding?
Human learning is a recurrent process of proposing hypotheses and trying them to gain knowledge. But not all possible hypotheses are equally meaningful for problem-solving. Understanding drives the progress of learning as it provides hypotheses that are both testable and useful.
Understanding develops within a subject domain, which is a system of hypotheses dependent on each other and proven in practice. Some of the dependencies between the hypotheses can be found independent of the language used to describe the subject. These structural patterns disclosed and followed by intelligence when making new hypotheses constitute understanding.
Machine understanding relates to transfer learning, also known as learning to learn.
The Training Method
The training method comprises 4 principles the development of machine understanding is based on:
- Narrowing the search area
- Training on examples
- Nesting in problems and solutions
- Reusing language structures
The principles can be used separately or in combinations.
A program example that trains AI to understand arithmetics shows the above principles at work. The supporting materials:
- light up the role of each principle in the training process
- explain how the principles can be realized in the training programs
- give an operational definition to the concept of machine understanding as a characteristic of learning behavior
The training examples can serve to spread the use of machine understanding into different subject domains.
Elephant for AI Developers
Machine understanding is double-sided: one side applies to training, the other regards learning. This is why Elephant principles need implementation in both training programs and learning agents to make an effect.
Training programs that necessarily count on machine understanding can be used to benchmark different AI technologies.