## The Backpropagation Debate: How the Brain Might Learn Like AI
Geoffrey Hinton, the celebrated pioneer of artificial intelligence, has been fascinated by the possibility that the human brain might employ a powerful learning algorithm known as backpropagation, similar to the one that powers advanced AI models like ChatGPT. This algorithm, often referred to as ‘backprop,’ has played a pivotal role in the remarkable rise of AI, enabling machines to perform tasks like writing coherent text, diagnosing illnesses from medical scans, and even navigating self-driving cars. However, for Hinton, creating better AI models was never the ultimate goal. His true ambition was to use these models as a lens to understand how the brain’s own neural networks solve complex problems.
The human brain learns through a subtle rewiring process, where connections between neurons, called synapses, are either strengthened or weakened. The brain contains billions of neurons, and understanding how it selects which synapses to adjust and by how much has been a major challenge for neuroscientists. Hinton’s contribution was to popularize backpropagation, a mathematical algorithm that efficiently solves this problem in artificial neural networks. But for a long time, it was considered too complex to have evolved in the human brain.
The recent emergence of AI models capable of remarkably human-like abilities has sparked renewed interest in whether the brain might utilize a process akin to backpropagation. However, uncovering the secrets of the brain’s learning process is not a simple task. Most of our current understanding comes from experiments on small brain tissue samples or limited groups of neurons in a laboratory setting. It’s unclear whether these findings translate to the intricate functioning of a living, learning brain.
Even with modern experimental techniques, where researchers can track hundreds of neurons simultaneously in living animals, reverse-engineering the brain’s mechanisms remains challenging. One dominant theory is Hebbian learning, which posits that neurons that fire together become more strongly connected. This idea explains simple associations, like Pavlov’s dogs learning to salivate at the sound of a bell, but for complex tasks like language acquisition, it appears insufficient. Even with extensive training, artificial neural networks trained solely on Hebbian learning lag far behind human performance.
Today’s top AI models, however, are engineered differently. To understand how they work, imagine an artificial neural network designed to identify birds in images. This network would consist of thousands of synthetic neurons arranged in layers. Images are fed into the first layer, which transmits information about each pixel to the next layer through AI-equivalent synaptic connections. Neurons in this layer might use this information to detect lines or edges, passing these signals to the next layer, which could identify eyes or feet. This process continues until the signals reach the final layer, responsible for making the final determination: ‘bird’ or ‘not bird’.
Central to this learning process is the backpropagation-of-error algorithm, or backprop. If the network misclassifies an image of a bird as something else, it generates an error signal upon realizing its mistake. This error signal travels backward through the network, layer by layer, strengthening or weakening connections to minimize future errors. When the model encounters a similar image again, the adjusted connections will lead to a correct classification: ‘bird’.
Neuroscientists have long doubted that backpropagation could occur in the brain. In 1989, soon after Hinton and his colleagues demonstrated the effectiveness of backpropagation for training layered neural networks, Francis Crick, the Nobel laureate who co-discovered the structure of DNA, criticized the theory in the journal Nature. He argued that neural networks using backpropagation were biologically ‘unrealistic in almost every respect.’ For one, neurons primarily transmit information in a single direction. For backpropagation to work in the brain, a perfect mirror image of each neuronal network would be required to send the error signal backward. Additionally, artificial neurons communicate using signals of varying strengths, while biological neurons send signals of fixed strengths, a characteristic that backprop isn’t designed to handle.
Despite these challenges, the success of artificial neural networks has reignited interest in the possibility of backprop-like mechanisms in the brain. Promising experimental evidence suggests that individual neurons in the brains of mice might respond to unique error signals, a key ingredient of backprop-like algorithms long considered absent in living brains. Researchers working at the intersection of neuroscience and AI have also shown that small modifications to backprop can make it more biologically plausible.
One significant study demonstrated that the mirror-image network once thought necessary for learning doesn’t have to be an exact replica of the original, although learning would proceed more slowly for large networks. This makes the possibility of backprop in the brain less implausible. Other researchers have found ways to bypass the mirror network altogether. If artificial neural networks can be endowed with biologically realistic features, such as specialized neurons capable of integrating activity and error signals in different parts of the cell, then backpropagation could occur with a single set of neurons. Some researchers have also modified the backprop algorithm to process spikes instead of continuous signals.
While the backpropagation hypothesis is gaining traction, some researchers explore alternative theories. In a paper published in Nature Neuroscience, Yuhang Song and colleagues at Oxford University proposed a method that reverses the flow of information in backprop. In conventional backprop, error signals drive adjustments in synapses, which in turn influence neuronal activity. Song and his team suggested that the network could first modify neuronal activity and then adjust the synapses accordingly. They termed this approach ‘prospective configuration.’ When they tested prospective configuration in artificial neural networks, they found that these models learned in a more human-like way, more robustly and with less training, compared to models trained using backprop. They also observed a closer alignment between the network’s behavior and human performance on various tasks, such as learning to move a joystick in response to visual cues.
Despite the advances in research, all of these theories remain just that: theories. Designing experiments to definitively prove whether backpropagation or any other algorithm is at work in the brain is surprisingly difficult. Aran Nayebi and colleagues at Stanford University recognized this challenge and employed AI to address it. They trained over a thousand neural networks using four different learning algorithms to perform various tasks. They monitored each network during training, recording neuronal activity and synaptic connection strengths. They then trained a separate ‘meta-model’ to infer the learning algorithm from these recordings. The meta-model successfully identified which algorithm had been used based on recordings from a couple of hundred virtual neurons at different points during training. The researchers hope that this meta-model could be applied to similar recordings from a real brain.
Identifying the algorithm, or algorithms, that the brain employs for learning would be a significant breakthrough in neuroscience. Not only would it illuminate the inner workings of this complex organ, but it could also help scientists develop new AI-powered tools to explore specific neural processes. Whether this research would lead to more effective AI algorithms remains uncertain. For Hinton, at least, backpropagation is likely superior to whatever mechanisms the brain uses. This debate highlights the intricate and fascinating relationship between AI and neuroscience, with each field offering potential insights into the other.