Neural Networks & Cognitive Science: How AI and the Brain Inform Each Other

Where the Metaphor Came From

In 1943, neurophysiologist Warren McCulloch and logician Walter Pitts published a paper describing a mathematical model of how neurons fire. Their "McCulloch-Pitts neuron" was a highly simplified abstraction: a node that receives inputs, applies a threshold, and either fires or doesn't. Boolean logic implemented in biology.

Frank Rosenblatt's 1958 perceptron built on this — a learning machine that adjusted the weights of its connections based on whether its outputs were right or wrong. The word "neural" was chosen deliberately. The implication was not subtle: we are building something like a brain.

That framing stuck. Through every winter and summer of AI progress since, we've talked about "neural networks," "deep learning" (layers of abstraction, like cortical hierarchies), and "attention mechanisms" (echoing selective focus in cognition). The language of neuroscience became the language of AI — even as the underlying mathematics diverged substantially from anything in biology.

How Artificial Neurons Actually Work

An artificial neuron takes a set of numerical inputs, multiplies each by a learned weight, sums the result, passes it through an activation function (like ReLU or sigmoid), and produces an output. Billions of these nodes, organized in layers, constitute a modern deep learning model.

Learning happens through backpropagation: the error in the output is propagated backward through the network, and each weight is adjusted slightly in the direction that would have reduced the error. Repeat this millions of times with millions of examples, and the network learns to represent patterns.

This is elegant and it works. GPT-4, AlphaFold, Stable Diffusion — all are variations of this basic architecture. The question is how much this has to do with biological neurons, and the honest answer is: less than the name implies.

The Gap Between Metaphor and Biology

Biological neurons are vastly more complex than their artificial counterparts. A single human neuron isn't a simple weighted adder. It integrates signals across its dendritic tree in nonlinear ways, adjusts its firing based on local chemistry, and communicates through spike timing as well as rate. A biological synapse involves dozens of proteins, complex feedback dynamics, and plasticity mechanisms that operate across multiple timescales simultaneously.

Backpropagation — the workhorse of deep learning — has no known biological equivalent. The brain doesn't appear to propagate error signals backward through layers in the way gradient descent does. Hebbian learning ("neurons that fire together wire together") and spike-timing-dependent plasticity are real biological mechanisms, but they operate very differently from how artificial networks are trained.

The neocortex uses sparse representations: at any moment, only a small fraction of neurons are active. Modern neural networks are dense by comparison. The brain runs on about 20 watts. A large AI model inference run can consume thousands of watts. These aren't footnotes — they're fundamental differences in architecture, efficiency, and implementation.

What AI Has Learned from Cognitive Science

Despite the implementation gaps, cognitive science has provided AI with some of its most important architectural ideas.

The concept of hierarchical feature extraction — the idea that perception builds up from simple features to complex representations — came directly from Hubel and Wiesel's work on the visual cortex in the 1960s. Convolutional neural networks (CNNs), which revolutionized computer vision, implement this hierarchy explicitly: edges → textures → parts → objects.

Attention mechanisms, which allow a model to selectively weight parts of its input (and which now power transformer architectures and every major language model), were partly inspired by research on how human visual attention works — how we direct focus to relevant parts of a scene rather than processing everything uniformly.

Working memory architectures — systems like LSTM networks that maintain state across sequences — were influenced by cognitive science models of how humans process sequential information. The idea that memory systems have different timescales and capacities comes directly from psychology.

What Cognitive Science Has Learned from AI

The influence runs both directions, though it's less often acknowledged.

Building AI systems that attempt to do what brains do has clarified what exactly brains are doing. When early speech recognition systems struggled with context-dependent phoneme recognition, it highlighted how much human speech perception depends on top-down priors — we don't just hear sounds, we hear sounds in light of what we expect the speaker to say. Building systems that failed at this taught researchers to be precise about what "understanding speech" actually requires.

The success of deep learning at visual tasks has reignited debates in cognitive science about whether human visual processing is more feed-forward than previously thought, or whether the correspondence reveals something deeper about what kinds of computational strategies work for this problem class.

Large language models have forced cognitive scientists to sharpen their claims about what language understanding requires. Does GPT-4 "understand" language? The question used to seem easy to answer. Now it requires careful definition of understanding — which is clarifying, even if it's uncomfortable.

Where the Comparison Breaks Down

Modern AI systems require massive labeled datasets to learn. Humans acquire language with far less explicit instruction, learning from sparse, noisy, and unlabeled experience in ways that still aren't fully understood. This sample-efficiency gap is one of the most significant differences between biological and artificial intelligence, and closing it is an active research challenge.

AI systems are notoriously brittle outside their training distribution. A vision model that achieves superhuman accuracy on ImageNet can be fooled by tiny, imperceptible perturbations to an image — adversarial examples that no human would misclassify. The brain's robustness to noise, occlusion, and novel inputs far exceeds anything currently achievable in artificial systems.

And crucially: AI systems don't have bodies. Embodied cognition research suggests that much of what we consider intelligence is deeply coupled to having a body that acts in and is shaped by a physical environment. Whether purely symbolic or statistical systems can replicate this without embodiment is an open question — and an important one.

What This Means for Consciousness

The neural network metaphor has an uncomfortable implication: if artificial systems can approximate so many cognitive functions, what exactly would make them conscious or not conscious?

One answer is that consciousness is tied to specific biological mechanisms — the particular way biological neurons signal, the specific neurochemistry of the brain — and artificial systems, however functionally similar, simply don't have those mechanisms. This is a form of biological naturalism.

Another answer is that consciousness is functional: if a system processes information in the right way, integrates it appropriately, and produces behavior that reflects that processing, then it is conscious regardless of what substrate implements it. This is a form of functionalism — and it makes the question of AI consciousness deeply uncomfortable, because it doesn't rule it out.

Most working AI researchers set this question aside. But it doesn't go away. Every time we build a system that does something we thought required consciousness — understanding jokes, generating metaphors, expressing apparent preferences — we have to either revise our theory of consciousness or decide that the system is doing something that looks like those things without being those things. Neither option is entirely satisfying. That's where the field sits right now.