TLDR: This research paper investigates the fundamental computational abilities that distinguish human symbolic thought from other animal cognition. It proposes that beyond basic dynamic binding, two hierarchical integration capacities are essential: multi-place predicates (representing relations between multiple variable arguments) and structure mapping (computing correspondences between systems of relations). Through simulations using the LISA cognitive architecture, the study demonstrates that both these capacities are independently necessary for basic symbolic thought, neither being able to compensate for the other. The findings offer insights into the late evolution of symbolic thought and highlight a key difference between the rapid, few-shot generalization of biological intelligence and the data-intensive learning of current machine learning models.
Human intelligence stands apart from that of most other animals, allowing us to develop complex language, mathematics, science, and engineering. This unique capacity is often attributed to our ability for ‘symbolic thought’ – the power to combine basic elements into a vast array of expressions and act upon them. But what exactly makes our brains capable of this, and why did it take so long to evolve?
A recent research paper, From Basic Affordances to Symbolic Thought: A Computational Phylogenesis of Biological Intelligence, by John E. Hummel and Rachel F. Heaton, delves into this fundamental question. The authors propose that while ‘dynamic binding’ – the ability to combine neurons into groups on the fly – is necessary for symbolic thought, it is not sufficient. They hypothesize that two additional, crucial abilities are required: the integration of multiple role-bindings into ‘multi-place predicates’ and the integration of multiple correspondences into ‘structure mappings’.
Understanding the Core Concepts
Let’s break down these key ideas:
- Dynamic Binding: This is a foundational ability, observed even in insects and cats. It allows an organism to link a property to an object in real-time. For example, a cat can dynamically bind the concept ‘can-walk-on’ to a novel surface, even if it has never seen that specific surface before. This enables universal inductive generalization for basic perceptual affordances.
- Multi-Place Predicates: This refers to the ability to represent relations between two or more objects where all arguments are free to vary. For instance, understanding ‘smaller-than (toy, box)’ is a multi-place predicate because both ‘toy’ and ‘box’ can change. This is qualitatively different from a single-place predicate like ‘can-walk-on (surface)’, where ‘me’ (the walker) is a fixed, implicit argument. Multi-place predicates enable recursion, allowing for an unbounded number of expressions from a finite vocabulary, vastly expanding expressive power.
- Structure Mapping: This is the capacity to compute systematic correspondences between elements of different representational systems. It’s what allows us to use one system to reason about another. Think of solving a physics problem with an equation: you map elements of the problem onto the variables in the equation. This is crucial for analogical reasoning, language comprehension, and understanding schemas. Without it, a symbol system would be isolated and functionally useless.
The Simulation Approach
To test their hypothesis, Hummel and Heaton used a biologically-inspired neurocomputational model called LISA (Learning and Inference with Schemas and Analogies). They systematically modified LISA to create four different cognitive architectures, all capable of dynamic binding, but varying in their other abilities:
- Dynamic Binding Only (DBO): Capable of dynamic binding but neither multi-place relations nor structure mapping.
- Relations Only (RO): Dynamic binding + multi-place predicates, but no structure mapping.
- Mapping Only (MO): Dynamic binding + structure mapping, but no multi-place predicates.
- Relations and Mapping (R&M): Dynamic binding + both multi-place predicates and structure mapping.
These architectures were then tested on four types of tasks, each designed to require different combinations of these capacities, from simple perceptual affordances to complex symbolic reasoning, without relying on prior training examples or specific features.
Key Findings from the Simulations
The results were striking and consistent with the hypothesis:
- DBO Task (Basic Affordance): All four architectures succeeded, demonstrating that perceiving basic affordances only requires dynamic binding.
- RO Task (Open-Ended Relations): Only the RO and R&M architectures, those capable of representing multi-place relations, succeeded. This showed that structure mapping alone (MO) could not compensate for the lack of multi-place relations.
- MO Task (Memory for Correspondences): Only the MO and R&M architectures, those capable of structure mapping, succeeded. This demonstrated that multi-place relations alone (RO) could not compensate for the lack of structure mapping.
- R&M Task (Symbol Systems): Only the R&M architecture, possessing both multi-place relations and structure mapping, successfully solved this complex task, which involved understanding how to operate a novel coffee maker based on its relational structure.
These findings strongly suggest that multi-place predicates and structure mapping are independent, yet both necessary, cognitive capacities for basic symbolic thought. Neither can fully compensate for the absence of the other.
Also Read:
- Enhancing Cognitive Models with Vector-Symbolic Memory
- Structured Prediction: How Combining LLMs with Inference Boosts Accuracy
Implications for Biological and Artificial Intelligence
The paper offers profound insights into the evolution of intelligence. The independence of these two capacities may explain why symbolic thought appeared so late in evolutionary history – it likely required two distinct ‘mutations’ to occur in the same organism.
Furthermore, the research highlights a critical difference between biological intelligence and modern machine learning approaches, particularly large language models (LLMs). While LLMs achieve impressive feats, they often rely on vast amounts of training data and ‘entangled’ representations, where relational roles are not truly independent of their fillers. This can lead to struggles with statistically unusual bindings, as illustrated by LLMs generating nonsensical images when asked to depict objects in novel, complex spatial relationships.
Biological intelligence, by contrast, excels at rapid generalization from very few examples, even a single one. This efficiency stems from its ability to represent invariants and dynamically bind them, and crucially, to form multi-place predicates and compute structure mappings. The LISA model, making inferences from single examples, offers a bio-inspired blueprint for developing more resource-efficient and rapidly generalizing artificial intelligence systems.
In conclusion, this research provides a compelling computational account of the minimal ingredients for symbolic thought, suggesting that the human brain’s unique cognitive power is built upon the foundational capacities of dynamic binding, multi-place predicates, and structure mapping.
