Chalcogenides? Like Ovonics?[1] That led decades ago to fast-read, slow-write memory devices. It was a dead end in electronics, but it did work. Makes sense for something that is slowly reconfigurable, like an FPGA.
"This advancement could shift information processing from data centers to local devices, providing real-time responses and extending battery life for applications like robotic motion control and autonomous vehicles."
It could also be about the big bet that a decentralized system, in which (e.g., in robotics) movement is determined joint by joint as opposed to a top-down global plan, is more efficient.
(Edit: although I realize that of course there is a Big long-studied discipline about it - Multiagent Systems.)
mdp2021
The article is about hardware that facilitates the integration of Fuzzy Logic and Neural Networks (hardware that is sufficient for NNs, provided they are FL oriented).
But Fuzzy Logic and Neural Networks can be opposites, when FL is adopted to have the Experts specify determined fuzzy rules ("if hot, then operate cooling"), and NNs are systems to induce rules, oftentimes obscurely (as opposed to the "generally clear, though specifically uncertain" rules of FL).
I suppose the contextual match between FL and NNs is that the FL rules will still be explicit, determined by the Experts, but the categories on which FL operates - the "membership function generation" will be left to the NNs.
westurner
I've heard it said that, we could do quantum computational operations with analog electronic components except for the variance in component quality.
How do these fuzzy logic electronic components overcome the same challenges as analog electronic quantum computing?
The article describes performance on a visual CNN Convolutional Neural Network ML task.
Is quantum logic the only sufficient logic to describe systems with phase?
What is most production cost and operating cost efficient at this or similar ML tasks?
> The TPU is constructed with a systolic array architecture that allows parallel 2 bit integer multiply–accumulate operations. A five-layer convolutional neural network based on the TPU can perform MNIST image recognition with an accuracy of up to 88% for a power consumption of 295 µW. We use an optimized nanotube fabrication process [...] 1 TOPS/W/s
"This advancement could shift information processing from data centers to local devices, providing real-time responses and extending battery life for applications like robotic motion control and autonomous vehicles."
Now that's a big stretch.
[1] https://en.wikipedia.org/wiki/Energy_Conversion_Devices
What stretch, from hardware to function?
It could also be about the big bet that a decentralized system, in which (e.g., in robotics) movement is determined joint by joint as opposed to a top-down global plan, is more efficient.
(Edit: although I realize that of course there is a Big long-studied discipline about it - Multiagent Systems.)
But Fuzzy Logic and Neural Networks can be opposites, when FL is adopted to have the Experts specify determined fuzzy rules ("if hot, then operate cooling"), and NNs are systems to induce rules, oftentimes obscurely (as opposed to the "generally clear, though specifically uncertain" rules of FL).
I suppose the contextual match between FL and NNs is that the FL rules will still be explicit, determined by the Experts, but the categories on which FL operates - the "membership function generation" will be left to the NNs.
How do these fuzzy logic electronic components overcome the same challenges as analog electronic quantum computing?
The article describes performance on a visual CNN Convolutional Neural Network ML task.
Is quantum logic the only sufficient logic to describe systems with phase?
What is most production cost and operating cost efficient at this or similar ML tasks?
For reference, from https://www.hackerneue.com/item?id=41322088 re: "A carbon-nanotube-based tensor processing unit" (2024) https://www.nature.com/articles/s41928-024-01211-2 :
> The TPU is constructed with a systolic array architecture that allows parallel 2 bit integer multiply–accumulate operations. A five-layer convolutional neural network based on the TPU can perform MNIST image recognition with an accuracy of up to 88% for a power consumption of 295 µW. We use an optimized nanotube fabrication process [...] 1 TOPS/W/s