To what extent are symbolic artificial intelligence and non-symbolic artificial intelligence distinct? How is the relationship between them reflected in modern AI systems?

Submitted for PH3C5 (Philosophy of Computing and Artificial Intelligence) Coursework

It is often said that artificial intelligence (AI) research developed along two broad tracks: a symbolic tradition (“Good old fashioned AI” or “GOFAI”) aiming to capture reasoning in explicit logical form; and a non-symbolic (or connectionist) tradition seeking to emulate cognitive abilities through networks of numerically weighted connections. In this essay, I will trace the historical development of both approaches to highlight their conceptual and methodological differences, but argue that modern AI systems blur this binary, by leveraging non-symbolic methods to perform symbol-like reasoning.

Early AI research was inspired by formal logic. Beginning with the Dartmouth Conference in 1956, researchers like John McCarthy, Allen Newell and Herbert Simon sought to formalise human thought as a process of rule-following, in line with methods inherited from proof theory and symbolic computation. Their conviction was that cognition could be captured by a structure akin to a formal language: if one modelled enough rules, the system would simulate human-like thinking. This approach led to projects like the Logic Theorist, an automated theorem prover capable of proving some theorems from Russell and Whitehead’s Principia Mathematica. Central to this paradigm was the belief that mental representations were best treated as strings of symbols that could be manipulated according to syntactic rules. Thus, providing a complete specification of these rules could enable machines to “think humanly”.

This approach led to some successes, for instance in early AI programs such as SHRDLU, a natural-language understanding system that users could instruct to manipulate blocks of various shapes and colours by parsing commands (like “pick up a big red block”) into formal syntactic representations, translating these into logical formulas, and reasoning with them to execute actions in its virtual environment (Winograd, 1972). But the initial optimism that “the range of problems [AIs] can handle will be coextensive with the range to which the human mind has been applied” (Simon & Newell, 1958) was tempered by the realisation that symbolic methods struggle with scaling complexity (Russell & Norvig, 2021, p. 39). The so-called combinatorial explosion, recognised in (Lighthill, 1973), refers to the massive increase in possible rule combinations and configurations as the number of variables increases, rendering brute-force symbolic reasoning infeasible for many real-world problems due to exponential computational demands. This heralded the first AI winter, a period of slow progress in the field, with the British government reluctant to fund efforts that were going nowhere. In particular, there seemed little hope of resolving the frame problem – “the challenge of representing the effects of action in logic without having to represent explicitly a large number of intuitively obvious non-effects” (Shanahan, 2016).

These limitations motivated the search for alternative methods. If brute-force symbolic reasoning failed to scale, then perhaps an adaptive system that learned from data could provide a solution. Whereas the symbolic approach sought to encode rules and declarative propositions explicitly, the idea emerged that a system might be able to learn such rules from examples. This led to what is now known as machine learning, where algorithms derive patterns from data rather than relying on hand-crafted knowledge. Instead of enumerating, for instance, every feature of a product for a system to classify different types of products, a learning system could observe many products and discover more generalisable rules for identifying different types of products. This learning process takes two primary forms: supervised learning, where models are trained on labelled data to predict correct outputs for unseen inputs; and unsupervised learning, which discovers patterns or clusters in unlabelled data to structure it in a meaningful way.

In non-symbolic approaches, exemplified by neural networks, the key insight is that intelligence need not reside in explicitly stored logical formulas. Early pioneers like (McCulloch & Pitts, 1943) proposed that neurons could be modelled mathematically; this foundational work was extended by (Hebb, 1949), whose rule that “neurons that fire together, wire together” introduced the idea of learning through adjustments to connection strengths. This is why the non-symbolic approach is also called the ‘connectionist’ approach: knowledge and understanding are held to be implicit in the patterned interactions of neuron-like units, each doing little more than summing inputs and producing an output signal. Learning occurs by iteratively adjusting these connection weights in response to feedback, rather than by modifying explicit rule sets. Over many such adjustments, the network converges to a set of weights that can classify objects into finite categories, or predict outputs from inputs in a way that best fits observed numerical data – without storing any identifiable “if–then” statements.

There are two important conceptual shifts here. The first concerns the nature of cognition: the symbolic tradition was aligned with the idea that thinking is internal language use, and its practitioners attempted to cause machines to think as humans think by formally encoding the laws of thought; but the connectionist approach suggests at least part of cognition happens below the threshold of conceptual-level expression. The second major change is in what researchers are trying to accomplish. GOFAI sought to replicate human thinking processes, driven by the belief that to achieve intelligence, machines must mirror the logical structures of human reasoning. Connectionist methods, however, focus not on replicating the processes of human thought but on producing intelligent behaviour – shifting the goal to cause machines to act as humans act. To understand this, it is worth noting that connectionist AI researchers are not trying to replicate the brain’s mechanisms. While initial machine learning systems such as the perceptron (the first neural model capable of learning simple classifications from data) were based on principles underlying brain function (Rosenblatt, 1958), these were plagued with issues – in particular, “when they are multi-layered and thus sufficiently expressive to represent non-linear functions, they were very hard to train in practice” (Bringsjord, 2018). It was the rediscovery of the backpropagation algorithm and its application to many learning problems, disseminated in Parallel Distributed Processing (Rumelhart & McClelland, 1987), that revitalised the connectionist paradigm. Backpropagation is a “method for training multi-layered neural networks [that] can be translated into a sequence of repeated simple arithmetic operations on a large set of numbers”. This approach resolved earlier issues by leveraging the fact that computing hardware is better suited to handling many simple, independent operations than a small number of complex ones (Bringsjord, 2018). But as (Lillicrap, Santoro, Marris, Akerman, & Hinton, 2020) observe, the human brain does not perform backpropagation to learn. This underscores that neural networks are computational abstractions designed to solve problems. They may draw inspiration from, or share parallels with, the way evolution shaped the brain to enable humans to solve those problems; but this functional perspective prioritises achieving outcomes over mirroring the underlying mechanisms of human cognition.

These major differences may suggest the two approaches are fundamentally distinct – indeed, (Harnish, 2002, p. 289) notes, “some even see the move to connectionism as a kind of Kuhnian paradigm shift”. I wish to resist this temptation and suggest that connectionist AI systems should more properly be understood as a superset of symbolic systems. That is, although GOFAI researchers had the ambition of replicating human-like cognition, their approach failed because symbolic reasoning is only a part of human cognition. But since it is a part of what humans do, a machine learning system seeking to “act humanly” must learn how to replicate the kind of syllogistic, symbolic reasoning in which people doubtless engage. And the issues of combinatorial explosion and the frame problem that rendered GOFAI useless (except perhaps for some expert systems – a point to which I will return) have, in a sense, been solved by modern connectionist systems; so the line between “symbolic” and “non-symbolic” is not as rigid as it once seemed.

To elucidate this argument, it is instructive to contrast two prominent thought experiments in the philosophy of AI, the Turing Test and Searle’s Chinese Room argument. (Turing, 1950) considers the question, “Can machines think?”, but concludes this is “too meaningless to deserve discussion”. Instead, he delineates an “imitation game” wherein a judge is to converse with a human and a machine, each trying to convince the judge they are the human. If the judge cannot reliably identify the human, the machine passes the Turing Test. Thus the question of whether it ‘really thinks’ is deemed irrelevant; the salient question is whether its behaviour resembles human thought convincingly enough to appear indistinguishable from human intelligence. By contrast, (Searle, 1980) is concerned above all with actual understanding: he distinguishes “weak AI” as the claim that AI systems are tools for solving problems that might appear to understand, from “strong AI”, the claim that the systems really do possess understanding. His argument is against strong AI, and proceeds thus: he imagines himself locked in a room with boxes of Chinese symbols and a rulebook (in English, which he understands) for how to manipulate these symbols to produce appropriate responses to Chinese input. Searle, by following these syntactic rules, can produce apparently fluent responses – “pass the Turing Test for understanding Chinese” (Searle, 1999) – despite not understanding Chinese. This, he argues, demonstrates that while a system might simulate understanding by manipulating symbols according to rules, it does not comprehend their meaning; so syntactic rule-following cannot give rise to strong AI.

These recall to mind the earlier framing, due to (Russell & Norvig, 2021, pp. 20-21), that the aim of symbolic researchers was to create systems that think humanly, while the aim of connectionist researchers became to create systems that act humanly. Searle’s argument is that a symbolic AI system, even if it could convincingly mimic human intelligence, would only be simulating thinking, and not really thinking. One counterargument to this is that although he may not understand Chinese (just as the CPU of a computer doesn’t), the whole room – the combination of the database of symbols, the rulebook, his following of the rules – does (Cole, 2024). But from where, even in this process of rule following, could understanding emerge? (Searle, 1980) contends that a human could in principle “internalize […] the system” – memorise the rulebook and the database – and thus produce convincing Chinese himself, still without understanding its semantic content. This seems suspiciously implausible; showing the speciousness of these arguments exposes the underlying flaw in the thought experiment. (Aaronson, 2013, pp. 39-40) observes that for the Chinese Room to convincingly mimic an intelligent Chinese interlocutor in real time, “if each page of the rule book corresponded to one neuron of a native speaker’s brain”, then the rule book would be larger than the Earth, “its pages searchable by swarms of robots travelling at close to the speed of light”. If the standard moves from one real time “imitation game” to handling the infinite possible linguistic inputs a human speaker can comprehend and generate, this practical absurdity becomes an in-principle impossibility: no finite rulebook could explicitly specify responses to every possible Chinese sentence. Thus Searle fails in his aim of describing a system that succeeds in acting humanly without real understanding; that is, he fails to show the Turing Test is inadequate.

It is noteworthy that Turing and Searle both focus in their arguments on language. This is useful, because in linking thought and behaviour, language shows that the goals of thinking humanly and acting humanly are more closely connected than it initially appears (unlike in say, chess, where a machine might plausibly use exhaustive search to play optimal moves without understanding). Searle attempted to show that a machine could exhibit behaviour as if it understood, when it fact it didn’t; his failure to do so has immense philosophical significance. An actual Chinese Room must bypass the need to explicitly encode the rules undergirding the process by which humans generate language. This is precisely the kind of task at which connectionist systems excel, by allowing machines to learn these inarticulable rules from data, and store them implicitly in the connections between neurons. Modern Large Language Models (LLMs) are systems of this sort; they approximate human linguistic competence by capturing statistical regularities within language to predict the next token. LLMs like GPT-4 can pass the Turing Test (Mei, Xie, Yuan, & Jackson, 2024) and understand what they say in the same sense humans do – (Cole, 2024) notes, “[ChatGPT] is loquacious if asked what a hamburger is, Searle’s example of something a natural language program cannot understand.”

This shows there are certain problems – like creative language use – that connectionist systems can solve, but since they cannot be reduced to a set of explicit propositional rules, cannot be solved by symbolic systems. Many have argued the converse is also true – that connectionist methods, since they only use statistical methods in lieu of explicit symbolic reasoning, are limited in their ability to perform tasks requiring compositionality, rule-based reasoning, or understanding hierarchical structures. In the wake of the frame problem, GOFAI systems were consigned almost entirely to “expert systems” seeking to simulate experts in domains – such as medical diagnosis or parts of mathematics – that rely heavily on these skills, and explicit rule-based syllogistic reasoning. Proponents of this view – like Gary Marcus, who in The Algebraic Mind (Marcus, 2001) argues that purely connectionist models lack the innate structure to capture the rule-based reasoning human cognition often requires – have thus been sceptical of the possibility that connectionist systems of the kind pervading contemporary AI research can make progress in these domains (believing general intelligence must come from a fusion of the two approaches). One of Marcus’ co-thinkers, Noam Chomsky, doubted even that such systems could attain creative language use (Chomsky, 2012).

But the combination of advances in computing power, large datasets made available by the Internet, and breakthroughs in deep learning techniques – “machine learning using multiple layers of simple, adjustable computing elements” (Russell & Norvig, 2021, p. 44) – led to unprecedented success of machine learning algorithms, even in areas seemingly better suited to symbolic systems. While GPT-2 learned only how to do natural language tasks at which it would be expected to excel, like “question answering, machine translation, reading comprehension, and summarisation” (Radford, et al., 2019), GPT-3 and its successors demonstrate capabilities extending into areas traditionally associated with symbolic reasoning. For instance, they can solve arithmetic problems, write functional code to achieve specific tasks, and unscramble words or anagrams (Brown, et al., 2020) – tasks Marcus and his ilk believed required explicit symbolic processing. Although LLM performance on such tasks is not perfect, their performance on relevant benchmarks has so far consistently improved as the models are scaled up (Petruzzellis, Testolin, & Sperduti, 2024).

The increasing success of larger connectionist models in these tasks, which the likes of Marcus and Chomsky thought them incapable, suggests a convergence between symbolism and connectionism. The symbolic school assumed various cognitive capacities are represented as distinct modules in the brain, each governed by hard-coded rules tailored to its specific function. But trying to build AI on this model of cognition failed, because it neglects that much of human thought operates below the level of explicit reasoning. As (Harnish, 2002, p. 338) argues, human cognition spans three interconnected levels: conceptual/symbolic, where we deal with familiar cognitive units like words and concepts; subconceptual/subsymbolic, where distributed patterns of activation give rise to these symbolic structures; and the neural level, where these patterns are grounded in the brain’s physical processes. Connectionist systems, in seeking to imitate human action rather than abstractly replicating the processes of human thought, have begun to capture these layers of cognition more effectively. Learning from data, they have gradually come to implement internal symbol manipulation – not as an explicitly programmed feature, but as an emergent property of their training and architecture. To mimic human-like action, which often involves reasoning and decision-making informed by symbolic thought, connectionist systems have effectively developed the capacity to parse logical arguments, handle compositional structures, and engage in the kinds of expert-system tasks once thought to be the sole domain of symbolic AI.

Thus, the distinction between symbolic and connectionist systems begins to dissolve. Although they represent very different approaches to building AI – the former seeking to replicate explicit rule-based reasoning, the latter leveraging data-driven adaption to emulate behaviour – connectionism subsumes symbolic AI in practice. GOFAI researchers were not wrong to consider symbolic reasoning a crucial aspect of cognition; but it is only one aspect, and a more nuanced model of cognition suggests modern neural network architectures implement the capabilities for symbolic operations, without explicit rules or syllogisms being programmed into them. This ‘implementationist’ view lends credence to the scaling hypothesis – as connectionist models become larger and better trained, they will increasingly excel at cognitive tasks requiring symbolic reasoning, approaching arbitrary proficiency with sufficient scale.

Bibliography

Aaronson, S. (2013). Quantum Computing since Democritus. Cambridge University Press.

Bringsjord, S. (2018, July 12). Artificial Intelligence (Stanford Encyclopedia of Philosophy). Retrieved from Stanford Encyclopedia of Philosophy: https://plato.stanford.edu/entries/artificial-intelligence/

Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J., Dhariwal, P., . . . Amodei, D. (2020). Language Models are Few-Shot Learners. Advances in Neural Information Processing Systems 33 (NeurIPS 2020). Retrieved from https://arxiv.org/abs/2005.14165

Buckner, C. (2019). Deep learning: A philosophical introduction. Philosophy Compass, 14(10), 1-19.

Chomsky, N. (2012, November 1). Noam Chomsky on Where Artificial Intelligence Went Wrong. (Y. Katz, Interviewer) Retrieved from https://chomsky.info/20121101/

Cole, D. (2024, October 23). The Chinese Room Argument (Stanford Encyclopedia of Philosophy). Retrieved from Stanford Encyclopedia of Philosophy: https://plato.stanford.edu/entries/chinese-room/

Domingos, P. (2015). The master algorithm: How the quest for the ultimate learning machine will remake our world. Basic Books.

Fodor, J. A., & Pylyshyn, Z. W. (1988). Connectionism and cognitive architecture: A critical analysis. Cognition, 28(1-2), 3-71.

Harnish, R. M. (2002). Minds, Brains, Computers: A Historical Introduction to the Foundations of Cognitive Science. Oxford: Blackwell Publishers Ltd.

Hebb, D. O. (1949). The Organization of Behavior: A Neuropsychological Theory. New York: John Wiley & Sons, Inc.

Lighthill, J. (1973). Artificial Intelligence: a general survey. Science Research Council.

Lillicrap, T., Santoro, A., Marris, L., Akerman, C., & Hinton, G. (2020). Backpropagation and the brain. Nature Reviews Neuroscience, 335-346.

Luxburg, U. v., & Schölkopf, B. (2011). Statistical Learning Theory: Models, Concepts, and Results. In D. M. Gabbay, S. Hartmann, & J. Woods, Handbook of the History of Logic (Vol. 10, pp. 651-706). Elsevier.

Marcus, G. (2001). The Algebraic Mind: Integrating Connectionism and Cognitive Science. MIT Press.

Marcus, G. (2022, March 10). Deep Learning Is Hitting a Wall - Nautilus. Retrieved from Nautilus: https://nautil.us/deep-learning-is-hitting-a-wall-238440/

McCulloch, W. S., & Pitts, W. (1943). A Logical Calculus of the Ideas Immanent in Nervous Activitiy. Bulletin of Mathematical Biophysics, 115-133.

Mei, Q., Xie, Y., Yuan, W., & Jackson, M. O. (2024, January 1). A Turing Test: Are AI Chatbots Behaviorally Similar to Humans? Retrieved from arXiv: https://arxiv.org/abs/2312.00798

Petruzzellis, F., Testolin, A., & Sperduti, A. (2024). Assessing the Emergent Symbolic Reasoning Abilities of Llama Large Language Models. Artificial Neural Networks and Machine Learning – ICANN 2024 (pp. 266-276). Springer.

Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., & Sutskever, I. (2019). Language Models are Unsupervised Multitask Learners. OpenAI. Retrieved from https://cdn.openai.com/better-language-models/language_models_are_unsupervised_multitask_learners.pdf

Rosenblatt, F. (1958). The perceptron: A probabilistic model for information storage and organization in the brain. Psychological Review, 386-408.

Rumelhart, D. E., & McClelland, J. L. (1987). Parallel Distributed Processing. The MIT Press.

Russell, S., & Norvig, P. (2021). Artificial Intelligence: A Modern Approach. Pearson.

Searle, J. R. (1980). Minds, Brains, and Programs. Behavioral and Brain Sciences, 417-424.

Searle, J. R. (1999). Chinese Room Argument. In R. A. Wilson, & F. C. Keil, The MIT Encyclopedia of the Cognitive Sciences (pp. 115-116). The MIT Press.

Shanahan, M. (2016, February 8). The Frame Problem (Stanford Encyclopedia of Philosophy). Retrieved from Stanford Encyclopedia of Philosophy: https://plato.stanford.edu/entries/frame-problem/

Simon, H. A., & Newell, A. (1958). Heuristic Problem Solving: The Next Advance in Operations Research. Operations Research, 1-10.

Turing, A. (1950). Computing Machinery and Intelligence. Mind, 433-460.

Winograd, T. (1972). Understanding Natural Language. Cognitive Psychology, 1-191.

Wolfram, S. (2023). What Is ChatGPT Doing ... and Why Does It Work? Wolfram Research, Inc.

Result

Mark: 68% (High 2.1)

Feedback:

This is a good essay tracing the history of artificial intelligence in regard to the symbolic / non-symbolic distinction. It is well-researched and provides a lucid high-level overview which efficiently covers lot of ground. At the same time, there are a number of points where details are cited which strengthen your case -- e.g. about the infeasible size of the rule book which be needed to actually implement Searle's Chinese room.

Overall, the fact that essay reads more like an overview than a critical assessment. However a thesis is stated towards the end -- i.e. 'the line between “symbolic” and “non-symbolic” is not as rigid as it once seemed' and that there is a 'convergence between symbolism and connectionism'. As such, one way in which the essay could have been improved is by stating this more clearly at the beginning and then structuring your exposition to support it directly. In order to do this, it would also have been useful to constrain the scope a bit -- e.g. by relating a bit less history or omitting the discussion of the Turing test.