Symbolic AI and Connectionism: Two Ways to Build Intelligent Machines
Symbolic AI represents knowledge with explicit symbols and rules, often drawing on formal logic. Connectionism, inspired by the brain, uses networks of simple units that learn from data. The debate between them shaped AI, and neuro-symbolic approaches now seek to combine the strengths of both.
The Divide
A long-standing divide in artificial intelligence separates two paradigms. One, often called symbolic or classical AI, treats cognition as the manipulation of explicit symbols according to formal rules. The other, connectionism, models intelligence with artificial neural networks: large collections of simple units whose connections have adjustable strengths. The Stanford Encyclopedia of Philosophy's entry on connectionism defines these networks as simplified models of the brain, comprising units analogous to neurons and weights that represent the strength of synaptic connections (Stanford Encyclopedia of Philosophy, Connectionism).
Both approaches aim to explain how machines can exhibit intelligent behavior, but they start from different metaphors. Symbolic AI takes the digital computer as its model, while connectionism looks to the biological brain. The philosophical debate between the two camps has shaped much of cognitive science.
Symbolic AI
Symbolic AI rests on the idea that intelligence emerges from explicit, rule-based manipulation of symbols. John McCarthy, a pioneer in the field, proposed using ideas from philosophical logic to formalize commonsense reasoning, a program described in the Stanford Encyclopedia's entry on logic-based AI (Stanford Encyclopedia of Philosophy, Logic-Based Artificial Intelligence). Such systems encode knowledge as logical sentences and reason by applying deductive rules.
A related philosophical thesis is the language of thought hypothesis, which holds that thinking itself occurs in a mental language, often called Mentalese. According to this view, mental states like beliefs and desires are relations to sentences in that language, and compositionality explains how complex thoughts arise from simpler parts, as the Language of Thought entry explains (the Language of Thought entry).
These systems excel at tasks that require clear logical reasoning and explicit knowledge, such as theorem proving, expert systems, and planning. However, they often struggle with learning from raw data and with handling noisy or ambiguous inputs.
Connectionism
Connectionism emerged from work in the 1940s. Warren McCulloch and Walter Pitts showed that networks of simple, neuron-like units could perform logical computations, a result highlighted in the Internet Encyclopedia of Philosophy's article on connectionism (Internet Encyclopedia of Philosophy, Connectionism). After a period of diminished interest, the field revived in the 1980s with new learning algorithms.
Connectionist networks learn by adjusting the strengths of connections between units. A common supervised method is the generalized delta rule, or backpropagation, which propagates an error signal backward through the network to update weights. This allows the network to gradually reduce its output errors.
Connectionist models have shown skill in perceptual tasks such as face recognition and reading, and they can tolerate noisy inputs. Their weakness lies in explaining systematic reasoning and compositional thought, a point central to the debate with classicists.
The Debate and Neuro-symbolic Synthesis
The debate between connectionists and classicists turned on whether neural networks could account for the systematicity and productivity of thought. Classicists argued that thought is systematic because the ability to understand one sentence entails the ability to understand related sentences, and productive because thinkers can generate an unbounded number of thoughts. These arguments are surveyed in the Computational Mind entry (the Computational Mind entry).
In recent years, the divide has begun to close. Neuro-symbolic AI attempts to combine the pattern recognition of neural networks with the explicit reasoning of symbolic systems. IBM Research describes its work in this area as aiming to give machines common sense and explainability, while also learning more like humans do (IBM Research, Neuro-symbolic AI).
Such hybrid systems hold promise for problems that require both effective learning from data and transparent logical inference, such as answering questions about images or generating reliable explanations.
| Aspect | Symbolic AI | Connectionism |
|---|---|---|
| Knowledge representation | Explicit symbols and rules | Distributed patterns across units |
| Reasoning | Deductive, logical inference | Learned associations, pattern completion |
| Learning | Typically not learned; knowledge encoded by hand | Learns from data by adjusting weights |
| Strengths | Clear explainability, logical rigor | Reliable perception, tolerance of noise |
| Weaknesses | Brittle, hard to scale to real-world data | Lacks systematic compositionality |
| Example tasks | Theorem proving, expert systems | Face recognition, language modeling |
Frequently asked questions
What is the main difference between symbolic AI and connectionism?
Symbolic AI uses explicit symbols and logical rules to represent knowledge and reason, while connectionism uses neural networks that learn patterns from data.
Why did connectionism decline and revive?
Early connectionist models had limitations, but advances in learning algorithms, such as backpropagation, led to a revival in the 1980s.
What is neuro-symbolic AI?
Neuro-symbolic AI combines neural networks with symbolic reasoning to get both learning from data and explainable logic.
Which approach is better for which tasks?
Symbolic AI suits tasks requiring clear rules and explanations, while connectionism excels at perception and pattern recognition. Hybrid systems aim to handle both.