In this blog, I want to share my current understanding of AI2Reason, a field that I made up to hold my goal of building AI that’s capable of reasoning (a reasoning machine).
In this post, I attempt to answer two central questions: why we should build a reasoning machine and why it’s reasonable to build reasoning machine (partially) based on statistics-oriented methods like ML. The answer to these two questions can be serve motivation of building reasoning machine based on ML methods. ****I will talk much less about how to do so (because I don’t know), which may be included in future post.
To answer this big questions, firstly, the philosophically interpretation of different levels of reasoning ability are introduced. Then, different contexts people have been working on are summarized, with special emphasis on formal mathematical reasoning. Lastly, recent progress on such formal reasoning context is surveyed.
Note that all arguments presented here are personal opinion only. If you are interested in this subarea and/or willing to work with me, feel free to contact me!
Update history:
Reasoning is a necessary condition for intelligence. An agent with only associative pattern matching is not only prune-to-error but also impossible to generate new idea. Although reasoning in general is a broad and vaguely defined word, we may simply define it as the attempt to answer questions by thinking about the reasons [14], or more practically, algebraically (relation is more important than elements) manipulating previously acquired knowledge in order to answer a new question (question-driven dynamics).
The reasoning ability is the elephant in the room of AI now. Current machine learning methods, powered by deep neural networks, excels at extracting predictive patterns from loads of data and training signals. On the other hand, the reasoning part is less developed. Along with the great success of LLMs, its weakness in reasoning is more and more concerned and generally with negative results [12] [13]. The success people got by incorperating LLMs into symbolic systems, e.g. [10, Stanislas Polu, 2020], doesn’t invalidate these concerns, but simply due to the other strength of LLMs, in my opinion.
AI2Reason is a trial in the era of computation to approach the ultimate goal of reasoning by incorporating data-driven approaches, e.g. machine learning, in symbolic reasoning systems.
The map mind of Artificial Reasoning System, and more
In this session, we want to answer the two central questions in high level by borrowing concepts and results from philosophy, the guide of science. We first answer why it’s reasonable to build reasoning machine (partially) based on statistics-oriented methods like ML by discussing detailed characterization of reasoning ability. We then answer why we should build reasoning machine by showing why high-level reasoning ability is important to future AI development and how come we’re not there yet.
The first step of all above goals is to understand reasoning in general.
As we previously said, reasoning is the attempt to answer a question by thinking about reasons and thus producing knowledge (which is factive) to answer it. It’s important to notice that reasoning itself does not guarantee anything. Reasoning is not affecting the world, it’s affected by the world. Even with the best of reasoning you might still end up with a false belief and thus fail to have knowledge to answer the question [14]. Therefore, we’d like to define good reasoning as the sort of thinking most likely to give you knowledge rather than mere opinion.
Argument is a practical way to understand reasoning, since reasoning is mainly presented through arguments in different forms. Good arguments needs to have true premises (the fit of the premises with the world), good logic (the fit of the conclusion with the evidence), conversational relevance (the fit of the argument with the conversation), and clarity (that it must be possible to tell whether these three kinds of fit exist) [14]. We will, of course, focus on the “good logic” in this session.