Symbolic AI and Connectionism
Artificial intelligence has long been both a contributor to and a testing ground for cognitive science theories. Two broad paradigms — symbolic AI and connectionism — represent fundamentally different theories of how cognition is implemented, and their competition has shaped cognitive science since the 1980s.
Symbolic AI (also called the 'physical symbol system hypothesis' or 'Good Old-Fashioned AI') proposes that cognition is essentially symbol manipulation — the processing of discrete representations according to explicit rules. A thought represents the world in a symbolic code (like mental language, or 'Mentalese'); reasoning involves applying logical inference rules to transform symbolic representations; and intelligence is the capacity to manipulate these symbols appropriately. Logic-based AI systems, production rule systems (like expert systems), and early natural language processing systems were implementations of this paradigm. Symbolic AI is well-suited to formal reasoning tasks: a symbolic AI system can reliably deduce conclusions from premises, execute algorithms, and perform logical inference. Its weakness is handling the messy, probabilistic, context-dependent, perceptually-grounded nature of real-world human cognition.
Connectionism (parallel distributed processing, PDP) takes a radically different approach: cognition emerges from the activity of large networks of simple interconnected processing units (artificial neurons), with knowledge stored not in explicit symbols but in the pattern of connection weights across the network. Rumelhart, McClelland, and the PDP group's 1986 books were foundational, demonstrating that simple neural networks could learn linguistic rules (like past tense formation in English) through exposure to examples without any explicit encoding of rules — the network generalized from the training data. Connectionist models naturally exhibit properties of human cognition including graceful degradation (partial damage reduces performance gradually rather than catastrophically), generalization to novel inputs, and context sensitivity.
Modern deep learning — the foundation of contemporary AI — is a descendant of connectionism, using large multi-layer artificial neural networks trained on massive datasets. Large language models (LLMs) like GPT-4 demonstrate that impressive language processing, reasoning, and knowledge can emerge from the scaled-up connectionist approach. The relationship between LLMs and human cognition is actively debated in cognitive science: do LLMs process language in ways that are similar to or fundamentally different from human language processing? This question is generating a productive new interface between AI and cognitive science.