The short version
The spikes must flow!I'd love to announce a new paper with that title, but sadly the editors at Neuron changed it.
Still v happy this paper is out because there's a revolution taking place in spiking neural networks and I want everyone to know about it.
Two of the things that make the brain interesting are (a) it is intelligent, it lets us make sense of very complex, noisy sensory data, (b) neurons use this super weird method of communicating. Now, for the first time, we can train spiking networks that can do hard tasks.
This is a game changer! We can finally begin to answer questions about how the brain uses patterns of spikes to compute in real-world situations. This is the question that got me into neuroscience in the first place!
So what changed? Methods from ML let us train neural networks at much harder tasks than before, but this was limited to artificial NNs, not spiking. Over the last couple of years, a number of tricks have been found to make it work for general case spiking neurons.
The code is relatively easy to write, but it's still quite slow at the moment and can only be used for a few hundred neurons. But, this is changing rapidly and there are going to be exciting times ahead over the next few years.
You can also take a look at recordings of the talks this review paper is based on at the SNUFA 2020 playlist.
Visualizing a joint future of neuroscience and neuromorphic engineering
Abstract
Links
Related organisations
Categories
The short version
I'd love to announce a new paper with that title, but sadly the editors at Neuron changed it.
Still v happy this paper is out because there's a revolution taking place in spiking neural networks and I want everyone to know about it.
Two of the things that make the brain interesting are (a) it is intelligent, it lets us make sense of very complex, noisy sensory data, (b) neurons use this super weird method of communicating. Now, for the first time, we can train spiking networks that can do hard tasks.
This is a game changer! We can finally begin to answer questions about how the brain uses patterns of spikes to compute in real-world situations. This is the question that got me into neuroscience in the first place!
So what changed? Methods from ML let us train neural networks at much harder tasks than before, but this was limited to artificial NNs, not spiking. Over the last couple of years, a number of tricks have been found to make it work for general case spiking neurons.
The code is relatively easy to write, but it's still quite slow at the moment and can only be used for a few hundred neurons. But, this is changing rapidly and there are going to be exciting times ahead over the next few years.
You can also take a look at recordings of the talks this review paper is based on at the SNUFA 2020 playlist.