New local learning algorithm achieves 87.3% accuracy without backpropagation
No backprop, no labels, no global signals—just local rules and temporal order.
Martin Andrews’ new paper demonstrates a paradigm shift in how neural networks can learn: using only local synaptic rules and the temporal structure of natural sensory streams, it replicates gradient-based learning without backpropagation. The mechanism combines spike-timing-dependent plasticity (STDP⁺) for potentiation and homeostatic plasticity via granule-cell-like neurons. Remarkably, the system requires no global error signals, no weight transport, and no label information—only pre- and post-synaptic firing rates and local statistics.
On a synthetic clustering task designed to isolate temporal order effects, input sequence alone boosted cluster separation ratio (CSR) from 0.83 (random order) to 2.49 (temporally ordered), a roughly 3.5σ improvement. On temporally ordered MNIST, a two-layer network trained end-to-end with these local rules achieved 87.3% linear-probe accuracy. This suggests that backpropagation may not be necessary for real-world learning—temporal contiguity and local plasticity could be sufficient. The work opens the door to more biologically plausible and energy-efficient AI systems.
- Achieved 87.3% linear-probe accuracy on temporally ordered MNIST using only local synaptic rules (STDP⁺ + homeostatic plasticity).
- Cluster separation ratio improved from 0.83 (random) to 2.49 (ordered) on synthetic clustering task—a 3.5σ effect from temporal ordering alone.
- Requires no backpropagation, no global error signals, no weight transport, and no labels—only local firing rates and temporal contiguity.
Why It Matters
Could enable efficient, brain-like AI training without backprop, reducing energy use and hardware constraints.