The Neuro-Symbolic AI (NS) initiative at IBM aims to conceive a fundamental new methodology for AI, to address the gaps remaining between today's state-of-the-art and the full goals of AI, including AGI. In particular it is aimed at augmenting (and retaining) the strengths of statistical AI (machine learning) with the complementary capabilities of symbolic AI (knowledge and reasoning). It is aimed at a construction of new paradigms rather than superficial synthesis of existing paradigms, and revolution rather than evolution. The primary goals of NS are to demonstrate the capability to:

1. Solve much harder problems
2. Learn with dramatically less data, ultimately for a large number of tasks rather than one narrow task
3. Provide inherently understandable and controllable decisions and actions

NS research directly addresses long-standing obstacles including imperfect or incomplete knowledge, the difficulty of semantic parsing, and computational scaling. NS is oriented toward long-term science via a focused and sequentially constructive research program, with open and collaborative publishing, and periodic spinoff technologies, with a small selection of motivating use cases over time.

The primary ones currently include the pursuit of true natural language understanding via the proxy of question answering; automatic data science, programming, and mathematics; and financial trading/risk optimization as ways to showcase the fundamental principles being developed. Research in NS is inherently multi-disciplinary and includes (among many other things) work in learning theory and foundations, optimization and algorithms, knowledge representation and acquisition, logic and theorem proving, reinforcement learning, planning, and control, and multi-task/meta/transfer learning.

Neuro-symbolic AI toolkit (NSTK) provide links to all the efforts related to neuro-symbolic AI at IBM Research. Some repositories are grouped together according the meta-projects or pipelines they serve.

LNNs are a novel Neuro = symbolic framework designed to seamlessly provide key properties of both neural nets (learning) and symbolic logic (knowledge and reasoning).

• Every neuron has a meaning as a component of a formula in a weighted real-valued logic, yielding a highly interpretable disentangled representation.
• Inference is omnidirectional rather than focused on predefined target variables, and corresponds to logical reasoning, including classical first-order logic (FOL) theorem proving as a special case.
• The model is end-to-end differentiable, and learning minimizes a novel loss function capturing logical contradiction, yielding resilience to inconsistent knowledge.
• It also enables the open-world assumption by maintaining bounds on truth values which can have probabilistic semantics, yielding resilience to incomplete knowledge.

Citation:

@article{riegel2020logical,
  title={Logical neural networks},
  author={Riegel, Ryan and Gray, Alexander and Luus, Francois and Khan, Naweed and Makondo, Ndivhuwo and Akhalwaya, Ismail Yunus and Qian, Haifeng and Fagin, Ronald and Barahona, Francisco and Sharma, Udit and others},
  journal={arXiv preprint arXiv:2006.13155},
  year={2020}
}

Knowledge base question answering (KBQA) is a task where end-to-end deep learning techniques have faced significant challenges such as the need for semantic parsing, reasoning, and large training datasets. In this work, we demonstrate NSQA, which is a realization of a hybrid "neuro-symbolic" approach. This system integrates multiple, reusable modules that are trained specifically for their individual tasks (e.g.semantic parsing, entity linking, and relationship linking) and require minimal training for the overall KBQA task, therefore addressing the challenges of end-to-end deep learning systems for KBQA.

Citation:

@inproceedings{kapanipathi2021leveraging,
  title={Leveraging abstract meaning representation for knowledge base question answering},
  author={Kapanipathi, Pavan and Abdelaziz, Ibrahim and Ravishankar, Srinivas and Roukos, Salim and Gray, Alexander and Astudillo, Ram{\'o}n Fernandez and Chang, Maria and Cornelio, Cristina and Dana, Saswati and Fokoue-Nkoutche, Achille and others},
  booktitle={Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021},
  pages={3884--3894},
  year={2021}
}

Neuro-Symbolic Agent (NeSA) is a neuro-symbolic approach to sequential decision making in real-world problems that is quantitatively and qualitatively better than end-to-end deep learning methods. NeSA uses domain knowledge, commonsense knowledge, and reasoning to reduce the number of interactions with the environment for learning the policy. The policy is represented as Logical Neural Networks (LNN), this enables explainable decision making and co-learning of the policy where the human can guide the training. Unlike, end-to-end system each stage of the pipeline can be individually verified and examined. For more details of the pipeline

Citation:

@inproceedings{kimura-etal-2021-neuro,
    title = "Neuro-Symbolic Reinforcement Learning with First-Order Logic",
    author = "Kimura, Daiki  and  Ono, Masaki  and  Chaudhury, Subhajit  and  Kohita, Ryosuke  and  Wachi, Akifumi  and  Agravante, Don Joven  and  Tatsubori, Michiaki  and  Munawar, Asim  and  Gray, Alexander",
    booktitle = "Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing",
    month = nov,
    year = "2021",
    address = "Online and Punta Cana, Dominican Republic",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2021.emnlp-main.283",
    doi = "10.18653/v1/2021.emnlp-main.283",
    pages = "3505--3511",
}