2020 BRN Discussion, page-15253

Interesting article here about AI controlled drones.

https://spectrum.ieee.org/automaton/robotics/drones/ai-powered-drone-extreme-acrobatics

which led me to this paper here by some of the same people:
https://arxiv.org/pdf/2003.02790v1

In this work, we investigated the applicability of feed-forward SNNs to regress angular velocities in continuous-time. We showed that it is possible to train a spiking neuralnetwork to perform this task on par with artificial neuralnetworks. Thus, we can confirm that state-of-the-art SNNtraining procedures accurately address the temporal errorassignment problem for SNNs of the size as presented in thiswork. Experimental results further suggest that deeper SNNsmight perform significantly better, but there are a number ofobstacles ahead. Backpropagation-based approaches requirethat we unroll the SNN in time at high-resolution. Thisrequirement poses serious challenges for optimization onGPUs both in terms of memory consumption and FLOPS. Ithas been a long-standing research goal to address these issuesand we believe it to be crucial to unlock the full potential ofSNNs.