• Recent publication in Proceedings of the National Academy of Sciences, 2024 on Foundations of Reasoning with Uncertainty via Real-valued Logics. Shows a new general way to analyze a large class of reasoning systems, including many modern neuro-symbolic approaches.
• Recent 3rd Neuro-Symbolic AI Summer School (NSSS ‘24), Sep 4-6, 2024. Featured a curation of star speakers and participation by thousands of researchers, students, and data scientists. Register to access the videos (to be made available shortly).

The Centaur AI Institute was established to achieve safe/reliable AI, with the hypothesis that an emerging set of “neuro-symbolic” ideas which incorporate both learning and knowledge/reasoning can provide the path toward resolving many if not most of the key issues preventing the use of AI in cannot-fail use cases such as in medicine, law, defense, transportation, financial systems, or conversational systems for performing transactions. More broadly, these issues are also fundamentally at the root of achieving greater capability and generality in AI. An upshot of these ideas is that new human-AI systems (“centaur AI” or human-in-the-loop systems) can be built where humans provide and interact with symbolic knowledge and are essential to AI rather than dispensible. The questions are both important enough and difficult enough to necessitate the formation of a new entity primarily motivated by the public good rather than profits, and insulated from the short-term motivations, timelines, and instabilities of even the largest industrial labs. As almost all other non-profit organizations addressing safe AI appear to be policy-focused rather than technically oriented, and almost all organizations addressing safe AI technically appear to be for-profit and focused on current mainstream technologies, we seek to fill the void rather than pursue either of those avenues.

The Institute is a 501(c)(3) whose mission is to illuminate the technical path to safer and more capable AI through fundamental research and advancing its adoption through education throughout the AI ecosystem. The Centaur AI Institute is funded through government research grants/contracts (e.g. NSF, DARPA, etc) and charitable donations by individuals, foundations or other organizations interested in helping to advance safe/reliable/neuro-symbolic AI. In particular the Institute can coordinate complex, multidisciplinary efforts which are otherwise difficult to put together and maintain on a long term basis. Specific activities include:

1. Research

A. 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 consider 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:

1. Could neural network-like models understand? (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.
2. Could neural network-like models ever be general? (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.
3. Could neural network-like models be much more efficient?
   • 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.

B. Research events, toward connecting the community. Centaur AI organizes research events such as workshops for the exchange of research ideas and the building and strengthening of a community around neuro-symbolic AI/safe AI research.

• 3rd research event (Winter 2024): AAAI 2024 Workshop on Neuro-Symbolic Learning and Reasoning in the era of Large Language Models (NucLeaR): 1.5 days, full room
• 2nd research event (Winter 2023): IBM Neuro-Symbolic Workshop 2023: 5000+ registrations
• 1st research event (Winter 2022): IBM Neuro-Symbolic Workshop 2022: 1500+ registrations

C. Open-source software, toward accelerating the community. Centaur AI creates and contributes to open source implementations of neuro-symbolic/safe AI research ideas, to help accelerate research, and ultimately put working technology in the hands of practicing data scientists.

2. Education

A. Educational events and materials. One of the main obstacles to the development of neuro-symbolic ideas is the inherent difficulty in spanning two very large and very different intellectual traditions - that of statistical AI or machine learning (ML), and that of symbolic AI or knowledge representation and reasoning (KRR) – most researchers are trained in one side but not the other. We hope to help onboard new researchers into this pursuit by providing educational entry points for researchers from both sides, via events with short courses and panels. We will also develop educational materials including presentations, courses, and books.

• 2nd education event (Summer 2023): Neuro-Symbolic Summer School (NSSS) 2023: 3500+ registrations
• 1st educational event (Summer 2022): Neuro-Symbolic Summer School (NSSS) 2022: 6000+ registrations

B. Guiding students. To help students interested in this emerging area (and indirectly their advisors), we can provide guidance on open problems to tackle and the history of approaches in the literature, as well as bring students into existing collaborations or start new research projects and otherwise make professional connections.

• Students are particularly welcome to reach out for guidance and advice, which may range from formal (e.g. internships, PhD committee service) to informal and year-round
• In Summer 2024, the Institute is hosting two research interns:
  • Bowen Li, PhD student, Carnegie Mellon University Robotics Institute. Project: Neuro-symbolic task and motion planning.
  • Jaiden Lee, undergraduate, Georgia Institute of Technology. Project: Augmenting LLMs for fact checking.

C. Advising organizations. To accelerate the development of the neuro-symbolic AI ecosystem, we advise organizations of all kinds on how to utilize neuro-symbolic AI technologies, including non-profits, government agencies, startups, venture capital and private equity firms, and large enterprises. Like many non-profit organizations, the Institute has a wholly-owned for-profit entity (Centaur AI Corporation) for the subset of its activities which further its mission but may also further the commercial interest of a third party - for example, advising companies on the use of neuro-symbolic technologies to advance their business. Similar to a university, we anticipate in the future helping to incubate startups based on neuro-symbolic technology, leveraging our past entrepreneurial experience.

Centaur AI has its own staff, a large network of close collaborators in academia (including Stanford, Carnegie Mellon University, Harvard, Yale, Purdue, Univ Sao Paulo, etc) and elsewhere who can be funded by Centaur, and many collaborators with jobs at commercial entities who work on a volunteer basis. Centaur AI forms project teams spanning experts in:

• KRR: Logical reasoning, knowledge representation, planning, probabilistic graphical models, causality
• ML: Deep learning, broader machine learning, reinforcement learning, statistics, data science
• Computation: Algorithms, optimization, 10x programming
• Beyond: Information theory, linguistics, semantics, neuroscience

Centaur AI is headquartered in California but operates via remote collaboration internationally. Centaur’s events are generally held virtually, with the exception of those that may be held as part of existing conferences such as workshops at AAAI and NeSy.

Education

For those seeking to learn more about neuro-symbolic AI including a crash course on basic concepts and a curated set of the most recent developments, see our resources above.

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)

Projects:

• Qualifications needed: We have projects for preparedness levels as early as highly talented undergraduate STEM students, as long as you can program well. (We also have pure mathematical projects, but these are at the graduate level only.) Otherwise, just bring your best productivity and willingness to learn fast by doing, and the team will be happy to help you advance.
• Commitment needed: The absolute minimum is one full day (8-12 hours) every week reliably, in order for the team to be able to depend on your contribution and invest members’ valuable time in you. The rewards you will get out will be proportional to the work you put in.
• Availability of paid positions: Most work in the Institute is voluntary. The main reward is the ability to work with and co-publish with strong researchers on meaningful AI problems, and get experience for your CV or a letter of reference if desired. Paid internships and part-time/full-time positions are generally by invitation as funds become available.

Operational and educational team

For those who want to help the Institute execute its mission and bring ideas for how to do it better, we welcome new members to the Operations team, which plans and creates our:

• Events and education program (being expanded)
• Discussion forums
• Outreach
• Grant proposals
• Seminar series, online resources, sponsorship program (being developed)

How we work

For some insight about our culture/style, here are is an incomplete list of some aspects (generally picked up through experience in or observation of other organizations):

• “One team”: We maximize our ability to achieve great things by rowing in one direction and mutual trust. This includes generosity in terms of help and acknowledgement of others.
• Ethics: For example, we are clear about non-profit vs for-profit pursuits and carefully segregate them into different organizations; proper crediting and meritocratic operation is essential.
• Debate: We are trying to solve difficult problems, so the answers are unclear and require robust ego-free and impersonal debate, which we can do by trusting in our teammates that anything can be questioned.

Commercial

For ambitious startups and larger enterprises seeking to take advantage of the powerful technologies emerging from this next wave of AI, we have the industrial experience to help you learn more and assess whether your problem can be well addressed, and even help you execute a proof-of-concept project through to a deployed system.

Donations

For those who want to help the cause - bringing to life a new, better form of AI for humanity - your financial support will be greatly appreciated and recognized!

Contact

For any of these, contact Alexander Gray (alexander.gray “AT” centaurinstitute.org).