I'm working on AI for robotics. My goal is a cheap robot that can do all your household chores.
Attention all students! I'm looking for people with the same goal. We have the robot, we have the compute, let's make it work! I care about practical results only, no theoretical advantages. Run it on the robot and show that it works!
Spiking neural networks add a temporal dimension to classic artificial neural networks. Instead of one-off activations, they model the dynamics of neurons that can emit short action potentials (spikes) over time. Similar to the brain, only certain parts of the network are active at any one time. This behavior opens up new possibilities for recurrent architectures, alternative learning methods and particularly energy-efficient inference.
The human brain learns new concepts surprisingly quickly and can adapt flexibly to unknown tasks. In comparison, modern neural networks usually require huge amounts of data for training and are restricted to narrowly defined fields of application. This raises fundamental questions: Is it due to common architectures such as multilayer perceptrons or transformers? Are they perhaps too inflexible to generalize across different areas? Do we need completely new training approaches beyond backpropagation and gradient descent? Or do we actually already have the right tool with the existing methods, only one crucial part of the overall picture is still missing?
