How Brains Learn Without Rules: The Connectionist Picture

Imagine you’re trying to teach a computer to recognize cats. You could try to write a rule: “If something has pointy ears, whiskers, and makes ‘meow’ sounds, it’s a cat.” But then you’d have to add more rules for cats with one ear, cats with no whiskers, cats that are sleeping, cartoons of cats—the list never ends. There’s always an exception.

Now imagine a different approach. You show the computer thousands of pictures of cats and thousands of pictures of things that aren’t cats. You let it figure out, on its own, what patterns separate the two. It never learns a rule it could write down. It just gets better and better at guessing.

That second approach is what connectionism is all about. It’s a way of thinking about how brains (and machines) might work—not by following strict rules, but by building up patterns of connections between millions of tiny processing units. And it challenges some of our oldest ideas about what thinking really is.

The Brain as a Giant Web

Here’s the basic picture. Imagine a network of simple units—let’s call them “nodes”—that are connected to each other. Some nodes receive information from the outside world (like your eyes sending signals about light). Some nodes send information back out (like signals to your muscles). And in between are many layers of “hidden” nodes that do the real work.

Each connection between two nodes has a weight—a number that says how strongly one node influences the other. A positive weight means “excite your neighbor.” A negative weight means “quiet down, neighbor.” When a node gets signals from all the nodes connected to it, it adds them up (weighted by their strengths), and if the total is high enough, it fires off its own signal to the next layer.

This sounds simple, but here’s the key insight: with enough nodes and enough connections, these networks can learn to do incredibly complex things—without anyone programming a single rule.

Philosopher Andy Clark, who writes about how minds work, once described the brain this way: not as a machine running a program, but as a “pattern-completing device” that’s constantly guessing what comes next.

How a Network Learns

So how do these networks learn? The most famous method is called backpropagation, and it works like this:

You give the network an input (say, a picture of a face).
The network makes a guess (is this male or female?).
You tell the network how wrong it was—not just “wrong” but how much wrong and in which direction.
The network adjusts all its connection weights just a tiny bit, in the direction that would have made its guess more correct.
You repeat this thousands or even millions of times with different examples.

Eventually, the network gets good at the task. It doesn’t just memorize—it generalizes. Show it a face it has never seen before, and it will usually get the answer right.

In the 1980s, researchers built a network called NETtalk that learned to read English text out loud. At first it sounded like random noise. Then it babbled. Then it sounded like it was speaking English double-talk. By the end, it could pronounce words it had never seen before. The network had somehow discovered the patterns of English pronunciation—without ever being told a rule like “a ‘c’ before ‘e’ sounds like ‘s’.”

Another famous example: a network trained to form the past tense of English verbs. It started by learning irregular verbs (“go” → “went,” “come” → “came”). Then when it learned regular verbs, it went through a phase where it over-regularized—it said things like “broked” instead of “broke.” This is exactly what human children do. The network wasn’t following a rule about adding “-ed.” It was just picking up on statistical patterns. But somehow that produced the same mistakes kids make.

Two Big Questions

Connectionism raises two huge questions that philosophers still argue about.

First: Is thinking really just pattern-matching? Or do we need something more—like rules and symbols—to explain things like language and logic?

The classical view of the mind (which dominated for decades) says that thinking is like running a computer program. You have symbols (like words or mental images), and you have rules for manipulating those symbols. When you understand “John loves Mary,” you’re manipulating symbols for “John,” “loves,” and “Mary” according to grammatical rules.

Connectionists say: maybe not. Maybe what we call “rules” are just the surface description of a much messier, statistical process happening underneath. The network doesn’t have a symbol for “add -ed to make past tense.” It just has weights that happen to produce the right result.

Second: Can connectionist networks really handle the complexity of human thought? This is where things get technical—and interesting.

The Systematicity Challenge

In 1988, philosophers Jerry Fodor and Zenon Pylyshyn launched a famous attack on connectionism. They pointed out something about human thinking: it’s systematic. If you can understand “John loves Mary,” you can also understand “Mary loves John.” If you can think that a red cube is better than a green square, you can also think that a green cube is better than a red square. These abilities come in clusters. You never find someone who understands one sentence but not its mirror image.

Fodor and Pylyshyn argued that connectionist networks have no built-in reason to be systematic. You could train a network to recognize “John loves Mary” without it being able to recognize “Mary loves John.” And if you can get it to do both, that’s just a lucky accident of training, not something guaranteed by the architecture.

The classical view, by contrast, explains systematicity automatically. If your mind stores the symbols “John,” “Mary,” and “loves” separately, and your rules let you combine them in different orders, then understanding one sentence does guarantee you can understand its mirror image. Systematicity comes for free.

This debate has been going on for over thirty years, and nobody has clearly won. Some connectionists have built networks that do show systematicity—but critics say those networks either cheat (by building in hidden advantages) or still don’t fully match human flexibility.

Distributed Representations: A Different Way to Store Meaning

One of the most interesting ideas to come out of connectionism is distributed representation. Here’s the contrast:

In a classical computer, information is stored in specific locations. The letter “A” is stored in one memory cell, “B” in another. In the brain, this would be like having a “grandmother neuron” that fires only when you think about your grandmother.

But that’s probably not how brains work. What we find instead is that information is distributed across many neurons. Your thought of your grandmother isn’t in one place—it’s a pattern of activation spread across thousands of neurons. The same neurons that help you think about your grandmother also help you think about other things.

This sounds strange, but it has huge advantages. If a few neurons die (which happens all the time as we age), the pattern is still mostly there. The memory degrades gracefully instead of disappearing entirely. Also, the relationships between ideas are built into the patterns. Similar ideas produce similar patterns. The meaning of a thought isn’t stored separately from the thought itself—it’s right there in the shape of the pattern.

Some philosophers think this could solve an ancient puzzle: How do brain states have meaning? In a classical system, symbols are arbitrary. The word “dog” doesn’t look anything like a dog. But in a distributed representation, the meaning is partly carried by the pattern’s internal structure, by how it relates to other patterns.

What This Means for How We See Ourselves

The connectionist picture of the mind is both exciting and unsettling.

It’s exciting because it seems more realistic than the classical computer model. Brains are made of neurons, not of logic gates. And connectionist networks do many things that classical computers struggle with: recognizing patterns, handling noisy or incomplete information, and gracefully degrading when parts break.

It’s unsettling because it suggests that a lot of what we think about thinking might be wrong. We tend to describe our own minds using the language of rules, reasons, and symbols. “I decided to go left because I believed the store was that way.” But what if that’s just a story we tell ourselves afterward? What if underneath, our brains are just pattern-matchers—incredibly sophisticated ones, but pattern-matchers nonetheless?

Some philosophers, called eliminativists, think that connectionism might eventually replace our everyday “folk psychology” altogether—the way we talk about beliefs, desires, and plans. If the brain doesn’t actually work that way, maybe we should stop describing it that way. Other philosophers resist this. They think beliefs and desires are real, even if they’re implemented in a connectionist brain.

Where We Are Now

Connectionism hasn’t settled the big questions. The debate between connectionists and classicists is still very much alive. But the conversation has changed.

Many researchers now think that the answer isn’t all-or-nothing. Maybe the brain uses both kinds of processing—pattern-matching for some things, symbol-manipulation for others. Maybe the two approaches can be combined in “hybrid” systems.

In recent years, connectionist ideas have exploded into something called deep learning, which powers everything from voice recognition to self-driving cars. Those systems are connectionist through and through—massive networks of simple units, trained on oceans of data, learning patterns no human could code by hand.

And yet, even the best deep learning systems still struggle with things humans find easy: learning from just one example, understanding cause and effect, truly grasping the meaning of a sentence rather than just its statistical patterns.

So the philosophical question remains: What is thinking, really? Is it following rules? Is it recognizing patterns? Or is it something we haven’t even come up with yet?

Connectionism doesn’t have the final answer. But it has given us a radically different way to ask the question—and that might be its most important contribution.

Appendices

Key Terms

Term	What It Does in This Debate
Connectionism	A way of explaining thinking as patterns of activity across many simple connected units, rather than as symbol-manipulation
Weight	The strength of a connection between two nodes; learning is adjusting these weights
Backpropagation	A learning method where the network gradually adjusts its weights based on how wrong its output was
Distributed representation	Storing information as a pattern across many units, not in a single location
Systematicity	The fact that understanding one thought or sentence automatically goes with understanding others built from the same parts
Classical (or symbolic) approach	The view that thinking is like running a computer program on mental symbols
Eliminativism	The view that our everyday way of describing minds (beliefs, desires, etc.) might turn out to be false and need replacing

Key People

Jerry Fodor (1935–2017) — Philosopher who argued powerfully that the mind runs on a “language of thought” and that connectionism can’t explain basic features of human reasoning.
Zenon Pylyshyn (1937–2022) — Cognitive scientist who collaborated with Fodor on the famous attack on connectionism for failing to explain systematicity.
Andy Clark — Contemporary philosopher who thinks connectionist and predictive-coding models offer our best picture of how the brain works.
Paul Churchland — Philosopher who argues that connectionism supports eliminativism—the idea that “folk psychology” is false and should be replaced.

Things to Think About

If you taught a connectionist network to recognize “cats” and “dogs,” and it got very good at it, would you say it knows what a cat is? What’s missing, if anything?
The systematicity argument says connectionist networks don’t guarantee that understanding one sentence means you understand its mirror image. But do you think you would understand “Mary loves John” if you’d never heard the name “Mary” in the subject position before? How do you actually learn that kind of thing?
If we’re really just pattern-matchers, what happens to ideas like “good reasons” or “logical thinking”? Can a pattern-matching brain be rational? Or would that require something like rules?
Some people worry that seeing the mind as a pattern-matching machine takes away something important—like free will, or the specialness of human thought. Does it? Or does it just give us a more accurate picture of what we already are?

Where This Shows Up

Artificial Intelligence: Every “deep learning” system you’ve heard of—ChatGPT, image recognition, voice assistants—is built on connectionist principles. These systems are the practical result of the philosophical ideas in this article.
Neuroscience: Real brains are, of course, neural networks. Connectionist models help neuroscientists think about how groups of neurons might produce complex behavior.
Education: If learning is about adjusting connection weights through repeated exposure, that has implications for how we teach. Drill and practice might matter more than we thought—but so might variety of examples.
Your own thinking: The next time you catch yourself saying “I know it, I just can’t explain it,” you’re describing exactly the kind of knowledge that connectionist networks excel at—pattern-matching that doesn’t reduce to rules.