Home > Articles

🧠🗣️💻 From Frege to ChatGPT: Compositionality in Language, Cognition, and Deep Neural Networks

🤖 AI Summary

• 🧠 Compositionality is the ability to compose familiar constituents into novel, complex combinations.
• 📐 The compositionality principle states a complex expression’s meaning is determined by the meaning of its parts and their combination.
• 🗣️ Human creative cognitive behaviors, such as language, are traditionally explained by the necessity of postulating compositionality.
• 🖥️ Classical symbolic AI systems guarantee compositionality intrinsically, but standard Deep Neural Networks (DNNs) are not architecturally designed to embody this.
• ❓ The Fodor and Pylyshyn objection traditionally targeted neural networks, claiming they lacked compositionality and were therefore non-viable cognitive models.
• 💡 Compositionality is required to explain productivity (unbounded novel thought generation) and systematicity (the intrinsic connection between understanding related thoughts, such as “John loves Mary” and “Mary loves John”).
• 🚀 Modern DNNs’ impressive performance in language domains calls into question the previously hypothesized limitations of these models.
• 🎓 Metalearning, or learning to learn, provides a novel perspective on how Large Language Models (LLMs) can reproduce the behavioral signatures of compositionality.
• ✅ Recent findings suggest contemporary DNNs have plausibly surmounted the first horn of the compositionality challenge, namely the objection regarding their empirical inadequacy.

🤔 Evaluation

• 🆚 The paper argues that metalearning offers LLMs a path to reproduce compositional behavior, potentially surmounting the “empirical inadequacy” critique.
• 🔬 Conversely, other research suggests it is premature to claim modern architectures have overcome the Fodor and Pylyshyn limitations.
• ⛔ This contrasting view points out that certain test cases compatible with meta-learning still provoke transduction errors and non-systematic behaviors, even within the Lake and Baroni setup (Position: Fodor and Pylyshyn’s Legacy — Still No Human-like Systematic Compositionality in Neural Networks - OpenReview).
• 🧐 Philosopher David Chalmers offers a different counter-critique to the classical Fodor/Pylyshyn argument, claiming the argument is logically defective because it would prove that the human brain, which supports compositional semantics, is also incapable of doing so (Connectionism and compositionality: Why Fodor and Pylyshyn were wrong - David Chalmers).
• 🗺️ The ongoing debate highlights key topics to explore for a better understanding.
• ❓ Topics include the current status of the “mere implementation” objection, which questions whether compositional LLMs simply mimic classical symbolic systems.
• 🔎 Further exploration is needed to empirically distinguish true systematic generalization from statistically driven behavior that merely resembles compositionality.

❓ Frequently Asked Questions (FAQ)

🧩 Q: What is the fundamental principle of compositionality in language and AI?

💡 A: Compositionality is the principle that the meaning of a complex expression, such as a sentence, is determined solely by the meaning of its constituent parts and the structural way those parts are combined.

🚫 Q: What is the core challenge of compositionality that symbolic AI posed to neural networks?

⚖️ A: The compositionality challenge, famously articulated by Fodor and Pylyshyn, is a dilemma arguing that neural networks are either empirically incapable of exhibiting the systematic behavior compositionality requires, or they merely implement a classical symbolic architecture, offering no novel cognitive explanation.

💻 Q: How do modern Large Language Models (LLMs) demonstrate compositional behavior?

🧠 A: Recent research suggests that modern LLMs, particularly those leveraging metalearning or in-context learning, can successfully reproduce the behavioral signatures of compositionality, potentially overcoming the argument that these networks are empirically inadequate for modeling human cognition.

📚 Book Recommendations

↔️ Similar

• 📘 The Oxford Handbook of Compositionality: A comprehensive resource edited by Oxford Academic exploring compositionality across linguistics, philosophy, and computation.
• 🏢🤖 Compositional Intelligence: Architectural Typology Through Generative AI by Daniel Koehler (Routledge): Investigates how compositional principles in urban design and architecture intersect with Generative AI and LLMs.

🆚 Contrasting

• 🌐 Parallel Distributed Processing Explorations in the Microstructure of Cognition: Volume 1 Foundations by David E Rumelhart and James L McClelland: The foundational text for connectionism, which presents an alternative, non-symbolic view of cognition that the compositionality debate fundamentally addresses.
• ♟️ The Algebraic Mind Integrating Connectionism and Cognitive Science by Gary F Marcus: Argues that connectionist models require innate, symbolic structure to achieve systematic generalization, upholding the necessity of a classical-style approach against pure neural networks.

🎨 Creatively Related

• 🧠💻 The Computational Brain by Patricia S Churchland: Explores the philosophy of mind and computational neuroscience, providing a biologically grounded framework for understanding how the brain supports complex cognitive functions like language.
• ✍️ Rethinking Writing Instruction in the Age of AI A Universal Design for Learning Approach by Randy Laist (CAST Professional Publishing): Discusses the practical impact of LLMs on critical thinking and language education, relating the theoretical capabilities of AI to real-world applications in human communication.