The gist of how it works is that analog circuits model the ion channels in neurons. You "program" the chip by adjusting parameters controlling the activation pattern of the "neurons" and connect the neurons through virtual "synapses".
An important detail: This emulates pyramid cells on hardware, it is not a software simulation. However you can "program" how the cells are connected to eachother by "virtual adressing". This is basically a programmable neural network on a chip. Dedicated to emulating neural networks.
You might see something like this in the medical field to help people move a robot arm. (By connecting it to your brain.)
Or maybe (i'm not sure) in the future self-driving car to, very fastly categorize objects in its surrounding (pedestrians, cars, etc)
this is super cool. would it work for deep learning without the associated data centre cost/energy consumption? Wonder how much further they can miniaturise the circuit
3 comments
[ 4.9 ms ] story [ 23.3 ms ] threadhttp://www.stanford.edu/group/brainsinsilicon/neurogrid.html
http://www.stanford.edu/group/brainsinsilicon/goals.html
The gist of how it works is that analog circuits model the ion channels in neurons. You "program" the chip by adjusting parameters controlling the activation pattern of the "neurons" and connect the neurons through virtual "synapses".
You might see something like this in the medical field to help people move a robot arm. (By connecting it to your brain.)
Or maybe (i'm not sure) in the future self-driving car to, very fastly categorize objects in its surrounding (pedestrians, cars, etc)