The Brain vs. Artificial Neural Networks

How brain science inspired AI — from the perceptron to deep learning — and how today's artificial neural networks are both similar to and very different from real brains.

Inspired by neurons

Artificial intelligence has borrowed from neuroscience since the start. McCulloch and Pitts's 1943 model neuron led to Frank Rosenblatt's perceptron (1958), a machine that learned to classify simple patterns by adjusting connection weights. Hebb's learning rule, Hubel and Wiesel's hierarchy of visual feature detectors, and the idea of distributed representations all shaped AI.

Deep learning

The backpropagation algorithm, popularised in 1986 by David Rumelhart, Geoffrey Hinton and Ronald Williams, made it possible to train networks with many layers. Kunihiko Fukushima's neocognitron (1980) and Yann LeCun's convolutional networks, directly inspired by the visual cortex, led to the deep learning revolution from 2012, when such networks dramatically improved image recognition. Today's large language models use the transformer architecture (2017), much less directly brain-inspired. In 2024, John Hopfield and Geoffrey Hinton received the Nobel Prize in Physics for foundational work on artificial neural networks.

Surprising similarities

Deep networks trained on images develop internal features resembling those in the visual cortex — edge detectors in early layers, more complex shapes later — and their internal activity predicts neural responses in monkey and human visual areas better than earlier models. Language models' internal representations partly predict brain activity during language comprehension. Reinforcement learning algorithms match dopamine signals.

Big differences

Neurons are not simple units: real neurons spike, have complex dendrites, many cell types, neuromodulators and glia. Learning: backpropagation requires sending precise error signals backward through every layer, which brains are not known to do in the same way; how the brain solves this "credit assignment" problem is an open question. Data: children learn language from a tiny fraction of the text used to train large language models. Energy: the brain runs on about 20 watts; training a large AI model can consume many megawatt-hours. Embodiment: brains evolved to control bodies in a world, continually learning without forgetting.

A two-way street

Neuroscience offers AI ideas about efficiency, continual learning and robustness; AI offers neuroscience models and tools to analyse vast datasets — and a test bed for theories of how intelligence might work.