Timothy Lillicrap, Daniel Cownden, Douglas Tweed, and Colin Akerman published “Random synaptic feedback weights support error backpropagation for deep learning” in Nature Communications in November 2016 (Vol. 7, article 13276). The paper addressed a long-standing objection to backpropagation as a model of how the brain learns.
The objection is the weight transport problem. Standard backpropagation sends error signals backward through a network using the exact same weights, transposed, that were used in the forward pass. There is no known biological mechanism by which a synapse could read out and reuse the precise weight of a distant forward synapse, so neuroscientists largely doubted the brain could implement backprop.
Lillicrap and colleagues showed something surprising: the backward weights do not need to match the forward weights at all. If the error is fed back through fixed, random weights, the network still learns effectively, because during training the forward weights spontaneously evolve to align with the random feedback so that the approximate error signal points in a useful direction. They named this effect feedback alignment and demonstrated it on a range of tasks.
For a general reader, the importance is conceptual: it removed a major reason to think brains could not be doing something like gradient-based learning, and it opened a productive research line on biologically plausible alternatives to backpropagation that continues to shape the neuroscience-AI dialogue.