iiinova — where your digital self keeps evolving

Replies

William Nourse @wnourse.bsky.social · 10/07/2026

(8/11) But wait, if we change the firing rate of the WTA neurons, then that also changes the rotational speed! To address this, we add a readout population which receives 1:1 input from the WTA. Using the same modulatory mechanism as the speed control, we can get radial ctrl.

Left: schematic of the network with readout neurons added. Changing the stim rate into the readout neurons changes the radius, either uniformly (center) or spatially unique (right). Network size N=50 for butter-smooth curves

110

William Nourse @wnourse.bsky.social · 10/07/2026

(9/11) This output mapping also allows us to fit our trajectory dynamics to arbitrary polar shapes, using essentially a lookup-table since the modulatory parameters are fairly predictable/consistent.

Two funky polar shapes, rose curve on the right and a harmonically distorted curve on the left.

100

William Nourse @wnourse.bsky.social · 10/07/2026

(10/11) Lastly, why did we bother going through all of this trouble to make this as a SNN vs a state machine or RNN? Well, we can execute this on neuromorphic hardware! We ran our network at various sizes on SpiNNaker2, and saw 2-3 orders of magnitude speed improvements, with 3-4 orders less energy!

Comparing GPU, single and multi-threaded CPU, and SpiNNaker2 for executing our network. SpiNNaker2 does way, way better

100