“Our approach combines several breakthroughs. We can very accurately estimate the precise parameters that control any neurons behaviour with high certainty. We have created physical models of the hardware and demonstrated its ability to successfully mimic the behaviour of real living neurons. Our third breakthrough is the versatility of our model which allows for the inclusion of different types and functions of a range of complex mammalian neurons.”
“By configuring individual ion channels of solid-state neurons with parameters estimated from large-scale assimilation of electrophysiological recordings, we successfully transfer the complete dynamics of hippocampal and respiratory neurons in silico. The solid-state neurons are found to respond nearly identically to biological neurons under stimulation by a wide range of current injection protocols”.
The related paper is here (open access): https://www.nature.com/articles/s41467-019-13177-3
They replicated one CA1 hippocampal neuron without its apical dendrites (power consumption of 139 nW). The backpropagating action potential is also replicated, but they don’t talk about the underlying learning rules (are STDP rules implemented in silico?).
139 nW x 86 billion of those neurons = around 10 kW
That is a great advance compared to other architectures, but still 3 orders of magnitude from our brains which consume 20 W.
The next challenge would be to scale to multiple neurons respecting topological rules, as discussed elsewhere in the forum!
They used an analog electrical circuit, which is crazy!
In the past, I’ve looked into using analog electronic circuits for simulating neurons. Analog circuits are really good at modeling differential equations, and in real time. However they tend to be large, inaccurate, and difficult to design. This paper appears to have overcome those issues.