For decades, Geoffrey Hinton, dubbed the ‘Godfather of AI’, has dedicated his research to artificial neural networks. His ambition wasn’t just to create sophisticated AI models capable of writing, diagnosing, and even driving; it was to unravel the mysteries of human brain function. He believed that by understanding how artificial networks learn, we could gain insights into how our own brains operate.
One of the key questions in this quest is how our brains selectively strengthen or weaken connections between neurons, known as synapses, during learning. This complex process, involving billions of neurons, has puzzled scientists for years. Hinton popularized an elegant mathematical algorithm called backpropagation to address this challenge in artificial networks. However, it was widely believed that this algorithm was too intricate to have evolved in the human brain.
The tide is turning as AI models increasingly exhibit human-like abilities. Scientists are now questioning whether the brain might employ a similar mechanism after all. While research on learning in live brains is incredibly challenging, recent studies provide promising hints. A 2023 study demonstrated that individual neurons in mice seem to respond to unique error signals, a crucial element of backpropagation-like algorithms previously deemed absent in living brains.
Moreover, researchers have shown that backpropagation can be made more biologically plausible with minor adjustments. Studies have demonstrated that the mirror-image network initially considered essential for backpropagation doesn’t need to be an exact replica, making the concept more believable. Others have discovered ways to eliminate the need for a mirror network altogether. By incorporating biologically realistic features into artificial neural networks, backpropagation can occur within a single set of neurons.
However, there are alternative theories. A study published in *Nature Neuroscience* proposes a ‘prospective configuration’ where neurons first adjust their activity and then adapt their connections, unlike traditional backpropagation. This approach, when tested in artificial networks, exhibited more human-like learning characteristics.
Despite these promising findings, definitively proving whether backpropagation, or any other algorithm, is at play in the brain remains a significant challenge. To address this, researchers at Stanford University used AI to analyze the learning algorithms employed in thousands of virtual neural networks. By monitoring neuronal activity and synaptic strength during training, they trained a meta-model to identify the algorithm used in each network. Their findings suggest that this meta-model could potentially be used to decipher the learning algorithm used by real brains.
Unveiling the algorithm that drives brain learning would be a groundbreaking achievement in neuroscience. It would not only illuminate the workings of our most mysterious organ but could also pave the way for AI-powered tools to understand specific neural processes. While the potential for better AI algorithms remains unclear, Hinton remains convinced that backpropagation surpasses whatever mechanism the brain uses.
The exploration of backpropagation in the brain is a captivating journey at the intersection of AI and neuroscience. This quest to decipher the learning process in our brains holds the promise of revolutionizing our understanding of ourselves and unlocking the potential of AI.
