> For the complete documentation index, see [llms.txt](https://neurosymbolicai.gitbook.io/neuro-symbolic-ai-in-practice/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://neurosymbolicai.gitbook.io/neuro-symbolic-ai-in-practice/part-i-motivation/chapter-1/1-2-neurosymbolic-agenda.md). # 1.2 The Neurosymbolic Agenda One of the most cited and comprehensive recent treatments of the neurosymbolic agenda is the 2023 IEEE Intelligent Systems paper by Sheth, Roy, and Gaur (2023), "Neurosymbolic AI — Why, What, and How." The paper makes three interlocking arguments, each of which motivates this book. ## Why The *why* of neurosymbolic AI is a diagnosis of the limitations of both dominant paradigms in isolation. **Pure neural systems:** * Are *data-hungry*: they require enormous labeled datasets that are expensive, slow, and often impossible to collect in specialized domains (rare diseases, edge-case safety scenarios, classified environments). * Are *opaque*: gradient-based models distribute knowledge across billions of parameters in a form that resists human inspection or formal verification. * Are *brittle under distribution shift*: a model trained on one dataset generalizes poorly to out-of-distribution inputs — the very inputs most likely in real-world deployment. * *Hallucinate*: they generate confident, fluent, and factually wrong outputs with no internal mechanism to detect or prevent the error. **Pure symbolic systems:** * Require *hand-crafted knowledge*: encoding domain knowledge as logical rules or ontologies is laborious, expensive, and inevitably incomplete. * Are *brittle under noise*: formal reasoners require clean, precisely structured input. Real-world sensor data, natural language, and images are messy — symbolic systems fail when input does not conform to their schemas. * *Do not learn*: classical symbolic systems cannot improve from data without explicit programmer intervention. The synthesis is not a matter of aesthetics. It is a practical necessity for any system that must operate reliably in the real world.(Sheth et al., 2023) ## What Sheth et al. map the landscape of neurosymbolic architectures along a spectrum from *neural-dominant* to *symbolic-dominant* systems. At one extreme, neural networks handle all computation and symbolic knowledge appears only as a lightweight constraint layer. At the other extreme, a symbolic engine handles all reasoning and neural components serve as sub-routine approximators. In the middle — and most interesting for practitioners — are *hybrid* systems where both components operate as co-equals, exchanging structured information through a well-defined interface. The authors introduce the *Knowledge-Infused Learning* (KIL) framework as a practical "how" for practitioners.(Sheth et al., 2023) KIL distinguishes three integration points: * **Pre-training:** Symbolic knowledge (ontologies, knowledge graphs, domain rules) guides corpus construction and sampling. The neural model begins training with a knowledge-shaped prior. * **In-training:** Symbolic knowledge appears as regularization terms, logical constraints on the loss function, or structured supervision signals. The model is penalized for outputs that violate known domain rules. * **Post-training:** Symbolic constraints are applied at inference time, filtering or re-ranking neural outputs that violate formal requirements. This three-point framework maps directly onto the architectural taxonomy in Chapter 4. ## How — Applied Domains Sheth et al. ground their claims in three application domains: 1. **Clinical NLP and mental health:** Biomedical ontologies (SNOMED-CT, DSM-5 criteria) are integrated with transformer-based models to produce interpretable, clinically faithful risk assessments. The symbolic layer ensures outputs align with diagnostic criteria even when training data is sparse.(Sheth et al., 2023) 2. **Cybersecurity:** Threat intelligence ontologies are integrated with neural anomaly detectors. The symbolic layer provides explainability (why is this traffic flagged?) while the neural layer handles the high-dimensional, noisy stream of network events. 3. **Disaster management:** Crisis ontologies are combined with social media NLP to produce structured situational awareness reports, where symbolic validation ensures geographic and temporal coherence. The consistent pattern: **the symbolic layer provides correctness guarantees and interpretability; the neural layer provides perceptual power and scalability.** > **Next:** [§1.3 — The LLM-Modulo Framework](/neuro-symbolic-ai-in-practice/part-i-motivation/chapter-1/1-3-llm-modulo.md) translates this diagnosis into a concrete architectural design pattern for practitioners building planning-capable AI systems. *Reference:* Sheth, Amit, Kaushik Roy, and Manas Gaur. "Neurosymbolic Artificial Intelligence (Why, What, and How)." *IEEE Intelligent Systems* 38.3 (2023): 56–62. *** --- # Agent Instructions This documentation is published with GitBook. GitBook is the documentation platform designed so that both humans and AI agents can read, navigate, and reason over technical content effectively. Learn more at gitbook.com. ## Querying This Documentation If you need additional information that is not directly available in this page, you can query the documentation dynamically by asking a question. Perform an HTTP GET request on the current page URL with the `ask` query parameter, and the optional `goal` query parameter: ``` GET https://neurosymbolicai.gitbook.io/neuro-symbolic-ai-in-practice/part-i-motivation/chapter-1/1-2-neurosymbolic-agenda.md?ask=&goal= ``` `ask` is the immediate question: it should be specific, self-contained, and written in natural language. `goal` is optional and describes the broader end goal you are ultimately trying to accomplish on behalf of the user. GitBook uses it to tailor the answer towards what is most useful for that goal. The response will contain a direct answer to the question and relevant excerpts and sources from the documentation. Use this mechanism when the answer is not explicitly present in the current page, you need clarification or additional context, or you want to retrieve related documentation sections.