In this era of information overload, we interact with smart devices every day. Have you ever wondered how these seemingly intelligent creatures know to "take an umbrella because it's raining"? Behind this is a profound transformation in causal reasoning.
A group of researchers from renowned academic institutions such as Microsoft and MIT have jointly developed a groundbreaking machine learning training strategy. This strategy not only overcomes the shortcomings of large machine learning models in logical reasoning but also achieves significant progress through the following steps:
Unique training method: Researchers have adopted a novel training method that may differ from conventional machine learning training techniques.
Improved logical reasoning: Their method significantly enhances the logical reasoning capabilities of large models, addressing previously existing challenges.
Building training sets with causal relationships: The research team utilizes causal relationship models to construct training datasets, which can reveal the causal connections between variables and help train models capable of understanding the causal logic behind the data.
Teaching the model basic axioms: They directly impart logical and mathematical basic premises to the model to help it better perform logical reasoning.
Despite having only 67 million parameters, the Transformer models trained through this method can rival the reasoning capabilities of GPT-4.
Causal reasoning, which sounds like a philosopher's exclusive territory, has actually permeated every aspect of our lives. For artificial intelligence, mastering causal reasoning is like learning to explain the world with "because...so...". But AI is not born with this ability; they need to learn, and this learning process is the story this paper discusses.
Axiomatic training method:
Imagine you have a very intelligent student, but they know nothing about causal relationships in the world. How do you teach them? Researchers have come up with a method – axiomatic training. This is like giving AI a "causation manual," allowing it to learn how to identify and apply causal rules through this manual.
Researchers conducted experiments with transformer models and found that this training method is truly effective! AI not only learned to identify causal relationships on small-scale graphs but also applied this knowledge to larger graphs, even those they had never seen before.
This research contributes by providing a new method for AI to learn causal reasoning from passive data. This is like giving AI a new "way of thinking," enabling it to better understand and explain the world.
This research not only shows the possibility of AI learning causal reasoning but also opens a door to potential future application scenarios for AI. Perhaps in the near future, our smart assistants will not only answer questions but also tell us why things are the way they are.
Research paper link: https://arxiv.org/pdf/2407.07612v1
