Symbolic Behaviour in Artificial Intelligence (2102.03406v2)

Abstract: The ability to use symbols is the pinnacle of human intelligence, but has yet to be fully replicated in machines. Here we argue that the path towards symbolically fluent AI begins with a reinterpretation of what symbols are, how they come to exist, and how a system behaves when it uses them. We begin by offering an interpretation of symbols as entities whose meaning is established by convention. But crucially, something is a symbol only for those who demonstrably and actively participate in this convention. We then outline how this interpretation thematically unifies the behavioural traits humans exhibit when they use symbols. This motivates our proposal that the field place a greater emphasis on symbolic behaviour rather than particular computational mechanisms inspired by more restrictive interpretations of symbols. Finally, we suggest that AI research explore social and cultural engagement as a tool to develop the cognitive machinery necessary for symbolic behaviour to emerge. This approach will allow for AI to interpret something as symbolic on its own rather than simply manipulate things that are only symbols to human onlookers, and thus will ultimately lead to AI with more human-like symbolic fluency.

• The paper demonstrates that symbolic behavior in AI is defined by dynamic social and cultural interactions rather than mere syntactic manipulation.
• It introduces evaluation dimensions like receptiveness, constructiveness, and malleability to assess AI’s capacity for symbolic fluency.
• The study suggests that immersive, human-interactive learning environments are crucial for developing nuanced and meaningful AI symbol use.

Symbolic Behaviour in Artificial Intelligence

The paper "Symbolic Behaviour in Artificial Intelligence" by Santoro et al. explores the complexities of achieving symbolic fluency in AI systems, reflecting a paradigmatic shift from traditional symbolic AI approaches to a behavior-focused understanding of symbol use. The authors argue that the essence of symbolic behavior transcends the syntactic manipulation of symbols and explores how symbols acquire meaning through social and cultural interactions. This perspective challenges classical views where symbol manipulation is primarily syntactic, emphasizing the pragmatic, dynamic, and interaction-driven aspects of symbols.

Key Concepts and Evaluations

The authors introduce several dimensions to evaluate symbolic behavior in AI systems, namely, receptiveness, constructiveness, embeddedness, malleability, meaningfulness, and grading:

1. Receptiveness: Current AI models, particularly neural networks, already demonstrate a high degree of receptiveness to established conventions, as evidenced by their success in tasks like LLMing and image understanding.
2. Constructiveness: This involves the ability to create new conventions. Current AI models exhibit limited constructive capacity as they often require predefined meanings for new concepts and lack naturalistic innovation compared to human communicators.
3. Embeddedness: Symbolic systems must adapt within the broader context they are used. Contemporary AI systems capture embeddedness better within their continuous and relational learning frameworks.
4. Malleability: The paper highlights the need for AI to engage with the flexibility of symbol meanings, adapting them based on context, a trait not robustly evident in current AI techniques.
5. Meaningfulness: Human symbolic understanding goes beyond mechanical manipulation to the semantics driving the reasoning process. Current AI struggles with demonstrating meaningful reasoning, often lacking introspective comprehension of its operations.
6. Grading: Symbol use is not binary, and the skill set around symbols is developmental and varied, which must be considered in AI systems.

Implications for AI Development

The authors propose that symbolic fluency in AI will emerge when learning-based agents are immersed in human socio-cultural interactions. By interacting and coordinating with human symbolic behaviors, AI can achieve a deeper understanding of meaning in conventions. This position advocates for leveraging large-scale datasets of human behavior and employing interactive human feedback, emphasizing the necessity for an immersive learning environment for AI development.

Furthermore, these insights call for AI systems to engage with symbolic conventions as humans do, involving a complex interplay between subjective interpretation and conventional meanings. The paper suggests that future AI could significantly benefit from engaging in environments that simulate human societal settings, allowing for the development of nuanced symbolic behavior.

Future Prospects

The potential future of AI hinted by this research extends beyond rule or syntactic-based processing to systems capable of understanding and altering the semantics of their interactions. This approach aligns with broader trends in AI research that favor embodied and experiential learning over purely computational methods. The trajectory involves a convergence of philosophical insights on symbols with the technical developments in AI, focusing on adaptive behaviors driven by cultural and social exigencies.

The exploration underscores significant theoretical implications, potentially transforming how AI systems are conceived and developed. It suggests a pivot towards AI architectures that emphasize contextual learning and dynamic adaptation to build systems that genuinely engage with the versatile nature of human-like symbol usage. The integration of these principles could lead to more competent AI capable of nuanced and meaningful interactions, resembling the rich symbolic engagement inherent in human intelligence.