Neuro-Symbolic AI Research
Attacking open problems, toward safer and smarter AI. Centaur AI conducts research projects ultimately aimed at what it regards as the most pressing open problems of AI, with a particular emphasis on developing a family of techniques which can, among other things, achieve:
Human Auditability and Controllability (HAC): By virtue of a human readable representation, the ability for humans to debug, improve, and control models by providing domain and world knowledge and ethical or other behavioral constraints.
Learning With Less (LWL): By incorporating high-level knowledge as well as raw data, the ability to learn with (much) less raw data, power, and cost.
Out-Of-Distribution (OOD) generalization: By incorporating rigorous reasoning, the ability to extrapolate to new situations (very) far from what was seen in training data.
These properties can all be considered key aspects of safe AI, as well as of AI that can generally perform better.
Examples of current practical open problems that can be addressed by this work:
Non-factuality in current generative AI systems.
Implicit bias in training sets which are too large to be fully vettable.
Massive costs creating an effective oligopoly in AI systems.
Centaur AI research threads consider both long-term conceptual foundations needed to enable such systems as well as development of cutting-edge new systems for standard AI tasks such as classification (e.g. as is common in data science), sequence-to-sequence transduction (e.g. as is done by LLMs), and sequential decision making (e.g. as is done by robots). Current questions being investigated include:
-
Item description
-
(as opposed to mimic sequences of text)
Could such models reason? (reliably, deductively as well as probabilistically)
e.g. we showed recently how to analyze a broad class of reasoning systems, including forms underlying neural networks, in Fagin et al., PNAS, 2024.
Could such models represent abstract knowledge? (e.g. grammars/programs, or a world model)
e.g. we showed recently how to measure the compositionality (a key property for representing general knowledge) of any model, including neural networks, in Ram et al., International Joint Conference on Artificial Intelligence, 2024.
-
(in the sense of AGI or Artificial General Intelligence)
Could such models generalize far from training data? (i.e. not just mimic solutions they have already seen)
e.g. we showed how a sequence-to-sequence approach that can utilize knowledge can generalize out-of-sample dramatically better, in Klinger et al., arxiv 2023.
Could such models infer causes? (especially in ways that go beyond the limitations of current causal models)
e.g. we began exploring how certain probabilistic semantics can allow modeling of causality in more realistic partially identified and cyclic cases, in Cozman et al., International Journal of Approximate Reasoning, 2024.
-
Could learning fundamentally be done with orders of magnitude less data?
e.g. we have developed the first quantitative information theory incorporating logical semantics, showing dramatically better compression is possible, in Lastras et al., arXiv 2023, revised 2024.
Can transformer-like models learn with much less data?
e.g. we have recently shown theory and experiments describing why certain kinds of sparsity (cf. Bengio’s “consciousness prior”) can yield faster learning, under review.
Join the Research.
New collaborators (researchers in academia or industry or government, graduate and undergraduate students, data scientists, engineers) are always welcome! All research by the Institute is collaborative and open, i.e. it works toward publications and open source software, aimed to advance the AI research community at large (only work performed in the Centaur AI Corporation subsidiary may be proprietary and any such work will be clearly identified as such and performed under appropriate contract).
Research is currently organized according to the following working groups:
Semantic parsing and generation (neuro-symbolic NLP for precision understanding of text and generative AI without hallucinations)
Grammar learning and parsing (better algorithms for compositional generalization, inductive logic programming, and abstraction)
Theory and practice of reasoning (new foundations for correct and efficient reasoning)
Probabilistic and causal semantics (more powerful semantics including handling ignorance and causality with cycles and partial identification)
Sequential decision making (greater efficiency and safety via neuro-symbolic reinforcement learning, planning, and robotics)
Biomedical applications (greater accuracy and safety in diagnosis, clinical trials, and community health)