Local Synaptic Rules Can Implement SIGReg Gradient Without Backpropagation

A new study has found that two local learning mechanisms in the brain—specifically spike-timing-dependent plasticity (STDP⁺) and homeostatic plasticity via flashlight granule-cell-like neurons—can effectively mimic a SIGReg-like self-supervised learning goal without using backpropagation. This method does away with the need for gradient calculations, global error signals, weight transport, or any labeling, focusing instead on the firing rates before and after synapses, local firing statistics, and the timing of natural sensory inputs. In a clustering test, an organized input sequence boosted cluster separation to 2.49, while random input kept it at 0.83, resulting in nearly a threefold increase just from the order of inputs. A two-layer network also showed similar results with ordered MNIST data, suggesting that biological neural circuits might achieve complex learning through local rules, offering a potential alternative to backpropagation in neuromorphic computing.