4 ms·
Not any kind of EE, just a nerd: could this kind of CCD circuitry be used to model a neural network, but in analog-land?
by cmrdporcupine 2y ago
Not any kind of EE, just a nerd: could this kind of CCD circuitry be used to model a neural network, but in analog-land?
- bgnn 2y agoThere are similar methods to model neural networks in analog. Switched capacitor circuits like CCD are semi digital as they are time discrete. This makes them very interesting for analog computing. It kinda never got popular though, as capacitors are much larger than transistors, and you need an ADC and a DAC to interact with a digital computer. In fact you need a lot of them, and they consume even more power than the digital alternatives. The biggest use case for this is sensor interfaces where the signal is still analog (not passed through an ADC yet). Voice recognition is a typical example where analog neural networks are used to certain level of success, and now people are pushing for insge recognition but the architecture of a digital camera isn't compatible with that, so I don't see much happening there. Funny fact is these kind of circuits are used in analog portion of the chips to implement rather complex calibration/training loops (correlation, LMS optimization, pattern recognition etc) heavily since early 90s. There's a lot of analog computing happening in every SoC.
- kazinator 2y agoThe values of the nodes in a network could be represented by charged capacitors. The functions which combine the voltages from multiple nodes could be operational amplifiers, with a clock driving everything: telling the nodes to sample their input that is calculated from the current values of nodes.