Comparing the Main Approaches to the History of Computing and Artificial Intelligence

Symbolic rule engines, connectionist networks, and statistical deep learning solved intelligence differently, on different evidence, at different costs — and none of the three ever fully retired the others.

Symbolic rule engines, connectionist networks, and statistical deep learning are three different approaches to machine inference, tested here on the same bench rather than ranked as stages of one march of progress. — Image prompt and art direction by Brecht Corbeel; image generated to that direction.

Abstract

Popular histories of artificial intelligence usually tell a succession story: symbolic logic gave way to neural networks, which gave way to deep learning, each one an upgrade that made the last obsolete. The archival record does not support a clean succession. Rule-based expert systems, connectionist networks, and statistical deep learning are three distinct approaches to the same underlying problem — how a machine should represent and revise what it "knows" — and they trade off interpretability, data requirements, and generalization in different, non-comparable ways. This article traces the three approaches from McCulloch and Pitts' 1943 formal neuron through the Dartmouth proposal, the perceptron, expert systems such as MYCIN, the AI winters, backpropagation, and the 2012 deep-learning turn, holding each approach to the same three dimensions rather than crowning a winner.

A succession story that the record does not support

The history of artificial intelligence is usually told as a relay race: symbolic logic ran the first leg, connectionist neural networks took the baton when symbolic methods stalled, and statistical deep learning took it in turn when neural networks scaled past a threshold around 2012. Each runner retires when the next one starts. It is a clean story, and it is largely a retrospective simplification.

The archival record shows something closer to three approaches to the same underlying problem — how a machine should represent what it has been given and revise that representation when it is wrong — developed partly in parallel, tried against each other repeatedly, abandoned and revived on different schedules, and never fully retiring one another. Rule-based symbolic systems, connectionist networks, and modern statistical deep learning are comparable on real, checkable dimensions: how legible their internal reasoning is to a person, how much labeled data and compute they require, and how well they generalize beyond the examples or rules they were given. They are not comparable as a single ranked ladder, because the dimension on which one approach wins is often exactly the dimension on which it loses to a rival built for a different constraint.

This article works through that comparison historically rather than abstractly, because the tradeoffs were discovered by people building specific systems under specific funding pressures, not derived from first principles in a seminar room.

Formal neurons and formal logic arrive together

The two lineages that would later compete start almost simultaneously, in the same decade, out of the same wartime interest in control and computation. Warren McCulloch and Walter Pitts published a model of an idealized neuron as a binary logical element in 1943, showing that networks of such units could in principle compute any logical function expressible by finite combinations of “and,” “or,” and “not” [1]. This is the founding document of connectionism: intelligence as the emergent behavior of many simple, uniform, interconnected units.

Seven years later, Alan Turing reframed the question of machine intelligence away from a definition of “thinking” and toward an operational test — whether a human interrogator could distinguish a machine’s typed answers from a person’s — while also discussing, in the same paper, the possibility of a “child machine” that learns rather than one that is fully programmed in advance [2]. Turing’s paper is a fact of publication history, not evidence that either approach is correct; it is cited here because it is the paper both later camps claim as an ancestor, and each reads a different half of it as decisive.

The 1955 Dartmouth proposal, written by John McCarthy, Marvin Minsky, Nathaniel Rochester, and Claude Shannon to solicit funding for a 1956 summer workshop, is the document that coined the term “artificial intelligence” and set the field’s early symbolic orientation: the proposal’s central conjecture was that every aspect of learning or intelligence could in principle be so precisely described that a machine could simulate it, and the workshop’s actual research agenda — problem search, symbol manipulation, formal reasoning — was heavily symbolic [3]. This is a matter of institutional history: which approach got named, funded, and staffed first, not which approach was later shown to be more correct. Connectionist work continued in parallel, notably Frank Rosenblatt’s 1958 perceptron, a probabilistic model of learning by adjusting weighted connections from example data rather than by specifying rules in advance [4]. Both lineages therefore predate any working definition of “artificial intelligence” as an institutionally organized field, and both were competing for the same scarce research funding from their first decade onward.

Dimension one: interpretability

The clearest, most durable difference between the two early approaches is how legible their internal state is to a person trying to understand why the system produced a given answer.

A symbolic, rule-based system represents knowledge as an explicit set of named propositions and inference rules — “if the organism is gram-negative and the infection site is the blood, consider these antibiotics” — so a person can, in principle, read the rule that fired, trace it back through the chain of prior rules that established its premises, and dispute a specific step. MYCIN, the Stanford system for recommending antibiotic therapy built through the 1970s, is the best-documented case: it operated with several hundred hand-authored production rules and was designed from the outset to explain its own reasoning by displaying the rule trace on request, and controlled evaluations judged its recommendations comparable in appropriateness to those of Stanford’s own infectious-disease faculty as assessed by outside experts [5]. That traceability is a real, load-bearing property, not a marketing description: it is what let clinicians and system builders audit a specific wrong recommendation back to a specific wrong or missing rule.

A connectionist system built on distributed weights has no equivalent unit to point to. Rosenblatt’s perceptron and its descendants store what has been learned as a large set of continuously adjustable connection strengths spread across many units, none of which corresponds to a single named fact [4]. Backpropagation, formalized for multilayer networks by David Rumelhart, Geoffrey Hinton, and Ronald Williams in 1986, made it practical to train many such layers by propagating an error signal backward through the network and adjusting every weight a small amount in the direction that reduces it [6]. The resulting network can perform well without any single weight, or even any single layer, being individually interpretable as “the rule that handles gram-negative infections.” This is not a defect that better tooling straightforwardly removes; distributed representation is the mechanism that gives these networks their other properties, discussed below, and interpretability is what is traded away to get them.

Figure 1. Symbolic systems made their reasoning legible as an explicit chain of rules a person could read, argue with, and trace back to a named premise. — Image prompt and art direction by Brecht Corbeel; image generated to that direction.

Statistical deep learning, the approach that scaled after 2012, inherits and sharpens the connectionist tradeoff rather than reversing it. Deep networks with many layers, trained on large labeled datasets, are if anything less interpretable than the shallow networks of the 1980s, simply because there are more layers and more weights between input and output. Contemporary interpretability research is an active attempt to recover some legibility after the fact — probing what individual units respond to, tracing which training examples most influenced a given output — rather than a property built into the architecture from the start. LeCun, Bengio, and Hinton’s 2015 review of deep learning, written after the approach’s post-2012 resurgence, describes representation learning explicitly as learning many levels of representation directly from raw data with only very general-purpose learning procedures, which is precisely the property that makes any one representation hard to name in advance [8].

Dimension two: data and compute requirements

The second dimension separates the approaches just as sharply, and in the opposite direction from interpretability.

Symbolic expert systems required expert time rather than large datasets. MYCIN’s rule base was built over years of interviews between Stanford computer scientists and infectious-disease physicians, encoding what those physicians already knew as explicit rules; the bottleneck was called the “knowledge acquisition” problem precisely because eliciting and formalizing expert judgment did not scale the way collecting raw examples does [5]. A symbolic system did not need thousands of labeled cases; it needed a small number of domain experts and a large amount of careful transcription and testing.

Connectionist networks reverse this: they need comparatively little hand-authored domain knowledge but instead need labeled training examples, and their appetite for those examples grew with each subsequent wave. Rosenblatt’s original perceptron learned from simple labeled input patterns with a single trainable layer [4]; Rumelhart, Hinton, and Williams’s multilayer backpropagation extended learning to deeper networks but at the cost of needing enough examples and enough computation to propagate error signals through additional layers without the training procedure stalling [6]. Krizhevsky, Sutskever, and Hinton’s 2012 AlexNet made the scale of this requirement concrete: a deep convolutional network trained on roughly 1.2 million labeled images across a thousand categories, using two graphics processing units running for about a week, cut the best prior image-classification error rate substantially on the standard ImageNet benchmark [7]. That result is a specific, verifiable fact about one benchmark and one dataset, not a general claim that deep networks always need millions of examples — some later architectures reduce that requirement — but it marks the moment the field’s dominant approach became explicitly dependent on large labeled datasets and parallel hardware in a way MYCIN’s rule base never was.

Figure 2. Connectionist systems stored what they had learned as distributed adjustable weights rather than as a list of named rules, trading a readable chain of reasoning for a capacity to generalize from examples. — Image prompt and art direction by Brecht Corbeel; image generated to that direction.

Stanford’s 2025 AI Index documents where that trajectory has led institutionally: continued growth in training compute, continued cost pressure to make inference cheaper even as frontier training runs grow larger, and a widening gap between the resources available to a small number of well-funded labs and everyone else [10]. That is a description of the current state of one approach’s resource demands, not a prediction about where the demands will stabilize.

Both approaches went through winters — for related but distinct reasons

Neither symbolic AI nor connectionism progressed smoothly. Both experienced periods, retrospectively labeled “AI winters,” when funding and public expectation fell well below the pace of technical progress, but the proximate causes differed in a way the interpretability and data dimensions predict.

Early connectionist research suffered a well-documented setback in the late 1960s tied to formal limits on what single-layer perceptrons could represent — a network with one trainable layer cannot separate certain patterns no matter how it is trained, a limitation that dampened funding for network-based approaches for roughly a decade until multilayer training methods matured. Symbolic expert systems suffered their own contraction in the 1980s, for a different reason: rule bases like MYCIN’s were expensive to build and brittle at the edges of their authored knowledge, and scaling a knowledge-acquisition process that depends on expert interviews turned out not to be cheaper than collecting more data, even though it required less data. Both slowdowns were about a mismatch between an approach’s actual scaling costs and the expectations that had been set for it, not evidence that either approach’s core idea was wrong. This is a historical description, not a prediction about whether the current deep-learning approach will face an analogous slowdown; the AI Index reports continuing high investment and capability growth as of 2025, and a future contraction, if one occurs, would need its own specific proximate cause identified rather than assumed from the earlier pattern [10].

Figure 3. Both approaches went through a period in which funding and expectations fell faster than the underlying research problems were solved, a pattern historians call the AI winters rather than a single failure of either method. — Image prompt and art direction by Brecht Corbeel; image generated to that direction.

Dimension three: generalization

The third dimension is the hardest to compare cleanly, because “generalization” means something different for each approach, and overstating apparent parity across the three is itself a common error this article aims not to repeat.

A symbolic system generalizes only within the scope its rules were written to cover, by design: MYCIN could reason correctly about infections and organisms its rule base explicitly represented, and its behavior at the edges of that coverage was a direct, traceable consequence of missing or incomplete rules — a property that made its failures diagnosable even when they were not prevented [5]. This is a form of controlled, bounded generalization: reliable inside a specified domain, and predictably silent or wrong outside it.

A connectionist or deep-learning system generalizes statistically from the distribution of its training examples, which lets it handle inputs it was never explicitly told about, provided those inputs resemble its training distribution closely enough. AlexNet’s improvement on ImageNet is evidence of exactly this kind of generalization on one specific, well-defined image-classification task and dataset, not evidence of open-ended reasoning [7]. LeCun, Bengio, and Hinton’s 2015 review frames the strength of deep learning explicitly as automatic discovery of the representations needed for detection or classification from raw data, which is a different and broader claim than MYCIN’s bounded rule coverage, but it comes with a correspondingly different and less predictable failure mode: a deep network can fail confidently and silently on inputs that look superficially like its training distribution but differ from it in ways the network never learned to detect [8]. Neither bounded rule coverage nor statistical distributional generalization is strictly “better” generalization; they are different guarantees, useful for different kinds of deployment risk, and conflating them is one of the more common overstatements in popular accounts of this history.

Figure 4. Statistical deep learning traded interpretability further still for raw capacity, scaling accuracy on large labeled datasets and cheap parallel hardware rather than on hand-written rules or small hand-tuned weights. — Image prompt and art direction by Brecht Corbeel; image generated to that direction.

Institutions chose among the three for reasons that outlast the technical debate

By the time the Association for Computing Machinery awarded its 2019 Turing Award to Yoshua Bengio, Geoffrey Hinton, and Yann LeCun for conceptual and engineering breakthroughs that made deep neural networks a critical component of computing, the field’s institutional center of gravity had moved decisively toward the connectionist and statistical lineage, a fact of professional recognition rather than a resolved technical argument about interpretability or data efficiency [9]. That institutional shift followed measurable results — the 2012 ImageNet result and its successors — but the underlying tradeoffs identified above did not disappear when funding moved: they are still the reason symbolic, rule-based reasoning persists today in domains such as tax software, some clinical decision support, and formal verification, where traceability and bounded, auditable behavior matter more than raw pattern-matching accuracy, while statistical deep learning dominates domains such as image recognition and language processing, where the training distribution can be made to resemble deployment conditions closely enough that distributional generalization is an acceptable risk.

Figure 5. No single dimension — traceability, data appetite, or raw generalization — makes one approach strictly better than the others; each was the right tool for a different constraint, and institutions chose among them accordingly. — Image prompt and art direction by Brecht Corbeel; image generated to that direction.

The comparison this history supports is therefore conditional rather than a ranking. Symbolic rule-based systems offer legible, auditable, bounded reasoning built from comparatively little data but a great deal of costly expert elicitation. Connectionist and statistical deep-learning systems offer broad, flexible pattern generalization built from comparatively little hand-authored domain knowledge but a great deal of costly labeled data and compute, at the price of an internal representation that resists inspection. Where an application needs a specific, traceable justification for a specific decision, the symbolic tradeoff still wins on its own terms. Where an application needs to handle inputs no one anticipated in detail, and can tolerate opaque, statistically-grounded failure modes, the connectionist tradeoff still wins on its own terms. The history of computing and artificial intelligence is not the story of one approach’s steady triumph over the others; it is the story of three different answers to the same design question, each kept alive wherever its particular tradeoff still fits the job.

What this history does not settle

A few claims are worth ruling out explicitly, because the succession narrative invites them by implication. It is not established that deep learning subsumes symbolic reasoning as a strict superset; a trained network that has never been given an explicit rule about drug interactions has no equivalent of MYCIN’s traceable rule chain to fall back on when it is uncertain, and post-hoc interpretability tools approximate that legibility rather than reproduce it exactly [8]. It is not established that expert systems failed because rule-based reasoning was a mistaken idea; the 1980s contraction in expert-systems funding is better explained, on the documented history of projects like MYCIN, as a mismatch between the cost of knowledge acquisition and the scale of coverage that was expected of it, a resource problem rather than a conceptual one [5]. And it is not established that current deep-learning scaling trends will continue indefinitely at the pace recorded in the 2025 AI Index; that report documents growth through 2024 and into 2025, not a law governing future years, and any claim about the next decade should be read as a scenario with a stated horizon and a disconfirmation condition, not as an extrapolated fact [10].

Where specialists in this history disagree — for instance, over how much credit for the 1980s connectionist revival belongs to backpropagation’s rediscovery versus to the parallel-hardware constraints of the period, or over how bounded MYCIN’s real clinical deployment actually was outside its evaluation studies — this article has tried to characterize the disagreement rather than resolve it in favor of whichever answer flatters one lineage’s later institutional success. The Turing Award citation for Bengio, Hinton, and LeCun recognizes conceptual and engineering breakthroughs in deep neural networks specifically, a professional judgment about impact within computing as a discipline, not an adjudication of the interpretability or data-efficiency tradeoffs discussed above [9]. Historical credit and technical superiority are different questions, and this comparison has deliberately kept them separate throughout.

Sources

1. Warren S. McCulloch and Walter Pitts. A logical calculus of the ideas immanent in nervous activity. Bulletin of Mathematical Biophysics (1943). DOI: 10.1007/BF02478259.
2. Alan M. Turing. Computing Machinery and Intelligence. Mind (1950). DOI: 10.1093/mind/LIX.236.433.
3. John McCarthy, Marvin L. Minsky, Nathaniel Rochester, and Claude E. Shannon. A Proposal for the Dartmouth Summer Research Project on Artificial Intelligence. AI Magazine (reprint of the 1955 proposal) (1955).
4. Frank Rosenblatt. The perceptron: A probabilistic model for information storage and organization in the brain. Psychological Review (1958). DOI: 10.1037/h0042519.
5. Bruce G. Buchanan and Edward H. Shortliffe (eds.). Rule-Based Expert Systems: The MYCIN Experiments of the Stanford Heuristic Programming Project. Addison-Wesley (Stanford Heuristic Programming Project) (1984).
6. David E. Rumelhart, Geoffrey E. Hinton, and Ronald J. Williams. Learning representations by back-propagating errors. Nature (1986). DOI: 10.1038/323533a0.
7. Alex Krizhevsky, Ilya Sutskever, and Geoffrey E. Hinton. ImageNet Classification with Deep Convolutional Neural Networks. Advances in Neural Information Processing Systems 25 (NeurIPS 2012) (2012).
8. Yann LeCun, Yoshua Bengio, and Geoffrey Hinton. Deep learning. Nature (2015). DOI: 10.1038/nature14539.
9. Association for Computing Machinery. Fathers of the Deep Learning Revolution Receive ACM A.M. Turing Award. ACM (2019).
10. AI Index Steering Committee. The 2025 AI Index Report. Stanford Institute for Human-Centered Artificial Intelligence (2025).

Originally published at https://absolutedigitalpublishers.com/articles/comparing-the-main-approaches-to-the-history-of-computing-and-artificial-intelligence.