02 — Research
Research
Brain-inspired spiking architectures for temporal sparse sensory processing.
-
2026 Learning without backprop: spiking networks for split learning Can a spiking network trained with a local rule learn useful features, and make split learning one-way? Poster
- What
- A spiking network that learns with a local rule (STDP), compared with a small backprop network on CIFAR-10.
- Why
- In split learning, backprop has to send gradients back to the client every step. A local rule doesn’t.
- So far
- About 72% accuracy without backprop, training in about five minutes, and far less data sent between client and server.
- Status
- In progress.
-
2025 Spiking networks on neuromorphic hardware How do accuracy, latency and energy change when a spiking network runs on a neuromorphic chip instead of a GPU? Poster
- What
- Running my spiking network on event-driven hardware and comparing it with a GPU.
- Why
- Spiking networks only save energy when the hardware is event-driven too.
- Plan
- Map the STDP model to a neuromorphic chip or simulator; measure accuracy, latency and energy per image.
-
2025 Synaptic performance of biological neurons What can real synapses (timing, plasticity, short-term dynamics) teach artificial spiking models? Poster
- What
- How biological synapses behave over time, and which of those behaviours spiking models leave out.
- Why
- Simplified STDP may be why deeper spiking layers learn so little.
- Plan
- Compare measured synaptic behaviour with the rules in my models, then test whether adding short-term dynamics changes what a layer learns.