The way I understand the article & paper they fix the in- and outputs of one network, electricity works out something akin to weights across their variable resistors, then they adjust the variable resistors in the other network based on those measurements. Rinse and repeat for your training set.
The way this works in detail is that the relative voltages across the different resistors/nodes on the first clamped network are the ones to get from a given input to a given output. It’s (at least for me) somewhat difficult to explain why this works out without a training aid circuit.
But essentially they know they voltage across all resistors for each element of the training set by asserting both in- and outputs on the first network, and then tweak resistor values in the second set to match those, which means without asserting the output value it still produces the desired output.
I wonder how this relates to the stuff Mythic0 has, which does similar stuff with analog weights in neural networks, though much more advanced. I have not found out yet how they train & manufacture their chips for a given training set.
Direct link to the paper: Demonstration of Decentralized, Physics-Driven Learning https://arxiv.org/abs/2108.00275
Seems like they had the desired voltages relative to ground for each node and tuned resistors until they got there.
How did they get the original model? AI?
The way I understand the article & paper they fix the in- and outputs of one network, electricity works out something akin to weights across their variable resistors, then they adjust the variable resistors in the other network based on those measurements. Rinse and repeat for your training set.
The way this works in detail is that the relative voltages across the different resistors/nodes on the first clamped network are the ones to get from a given input to a given output. It’s (at least for me) somewhat difficult to explain why this works out without a training aid circuit.
But essentially they know they voltage across all resistors for each element of the training set by asserting both in- and outputs on the first network, and then tweak resistor values in the second set to match those, which means without asserting the output value it still produces the desired output.
I wonder how this relates to the stuff Mythic0 has, which does similar stuff with analog weights in neural networks, though much more advanced. I have not found out yet how they train & manufacture their chips for a given training set.