3 ms·

Did you look at the paper or just throwing the analogy out there? It seems the described problems are integer programming which wouldn’t have gradients as in b

Did you look at the paper or just throwing the analogy out there? It seems the described problems are integer programming which wouldn’t have gradients as in back propagation. The magic seems to lie in the so called self organizing logic gates detailed https://arxiv.org/abs/1512.05064 https://arxiv.org/abs/1512.05064

Animats 9y ago
Right. See the PDF of the paper at [1]. They're mapping Boolean logic into a kind of analog system. "SOLGs can use any terminal simultaneously as input or output, i.e., signals can go in and out at the same time at any terminal resulting in a superposition of input and output signals ... The gate changes dynamically the outgoing components of the signals depending on the incoming components according to some rules aimed at satisfying the logic relations of the gate. ... A SOLG ... can have either stable configurations ... or unstable ... configurations. In the former case, the configuration of the signals at the terminals satisfies the logic relation required ... and the signals would then remain constant in time. Conversely, if the signals at a given time do not satisfy the logic relation, we have the unstable configuration: the SOLG drives the outgoing components of the signal to finally obtain a stable configuration." This is sort of like supervised training of a neural net. I think. Test cases with both inputs and desired outputs are needed, and applying them to the network pushes the parameters towards values that yield the desired outputs. It's kind of like deliberately over-training a neural net to encode some explicit function. The paper is vague about how you train this thing. It seems like it has to be driven by test cases, but they don't say much about that. It's not clear that this scales. It ought to work for small numbers of gates, but for large numbers, does this process converge? [1] https://arxiv.org/pdf/1512.05064.pdf https://arxiv.org/pdf/1512.05064.pdf

Animats 9y ago

Right. See the PDF of the paper at [1]. They're mapping Boolean logic into a kind of analog system. "SOLGs can use any terminal simultaneously as input or output, i.e., signals can go in and out at the same time at any terminal resulting in a superposition of input and output signals ... The gate changes dynamically the outgoing components of the signals depending on the incoming components according to some rules aimed at satisfying the logic relations of the gate. ... A SOLG ... can have either stable configurations ... or unstable ... configurations. In the former case, the configuration of the signals at the terminals satisfies the logic relation required ... and the signals would then remain constant in time. Conversely, if the signals at a given time do not satisfy the logic relation, we have the unstable configuration: the SOLG drives the outgoing components of the signal to finally obtain a stable configuration." This is sort of like supervised training of a neural net. I think. Test cases with both inputs and desired outputs are needed, and applying them to the network pushes the parameters towards values that yield the desired outputs. It's kind of like deliberately over-training a neural net to encode some explicit function. The paper is vague about how you train this thing. It seems like it has to be driven by test cases, but they don't say much about that. It's not clear that this scales. It ought to work for small numbers of gates, but for large numbers, does this process converge? [1] https://arxiv.org/pdf/1512.05064.pdf https://arxiv.org/pdf/1512.05064.pdf