Omer Faizan
← Research

Research poster · work in progress · 2026

Learning without backprop: spiking networks for split learning

Omer Faizan · M.S. Data Science, Montclair State University

1 Question

Split learning keeps raw data on the client, but backprop still sends gradients back every step. Can a network that learns with a local rule, no gradients at all, learn useful features, and how much traffic does that save?

2 Method

image whitendecorrelate STDP conv512 filterswinner-take-all poolregions ridgereadout local learning rule, no gradients

A spiking network trained with STDP, compared with a small backprop network (ResNet-8) on CIFAR-10, under the same split, augmentation and readout.

3 The split

clientdata stays herefirst layer serverreadout + restof the model features → (compressed) ← gradients (backprop only) spiking net: one-way traffic, nothing comes back

4 Results

~15 ptsabove the closest published STDP result
~8×less data sent in a real client/server split
2 h → 5 mintraining time, same accuracy

Test accuracy, CIFAR-10 (rounded)

Spiking netSpiking net (STDP): ~72% · no backprop~72%ResNet-8ResNet-8: ~85% · backprop~85%

Accuracy by epoch, same readout for both (approximate shape)

60%70%80%90%0246810epochResNet-8, epoch 0: ~60%ResNet-8, epoch 1: ~72%ResNet-8, epoch 2: ~76%ResNet-8, epoch 3: ~78%ResNet-8, epoch 4: ~80%ResNet-8, epoch 5: ~82%ResNet-8, epoch 6: ~82%ResNet-8, epoch 7: ~83%ResNet-8, epoch 8: ~84%ResNet-8, epoch 9: ~84%ResNet-8, epoch 10: ~84%ResNet-8Spiking net, epoch 0: ~60%Spiking net, epoch 1: ~72%Spiking net, epoch 2: ~72%Spiking net, epoch 3: ~72%Spiking net, epoch 4: ~72%Spiking net, epoch 5: ~72%Spiking net, epoch 6: ~72%Spiking net, epoch 7: ~72%Spiking net, epoch 8: ~72%Spiking net, epoch 9: ~72%Spiking net, epoch 10: ~72%Spiking net

Spiking net (STDP) ResNet-8 (backprop)

5 What matters inside

Accuracy lost when one part is removed (points, rounded)

Winner-take-allRemove winner-take-all: about −47 points−47Region poolingRemove region pooling: about −13 points−13WhiteningRemove whitening: about −12 points−12Adaptive thresholdsRemove adaptive thresholds: about −4 points−4Square rootRemove square root: about −1 points−1Lateral inhibitionRemove lateral inhibition: about −1 points−1Residual shortcutRemove residual shortcut: about −0 points−0

6 Takeaways

  • Competition (winner-take-all) is the whole game; without it the network collapses.
  • The STDP features are learned almost entirely in the first epoch, then plateau.
  • The gap to backprop is real, but the traffic is one-way and much smaller.