4 ms·
In 10 years major chip design houses might start seriously consider using analog computation for machine learning.
by p1esk 7y ago
In 10 years major chip design houses might start seriously consider using analog computation for machine learning.
- newyankee 7y agoCan you elaborate ?
- 0-_-0 7y agoMost networks currently are float32. Low accuracy networks are possible, and the random noise of an analog circuit is basically a regularizer for a neural network, instead of being unwanted or incorrect. It's a feature, not a bug!
- p1esk 7y agoAnalog computing is much more efficient than digital one. For example, you can perform a multiplication of two numbers using a single transistor (or memristor/etc), if you only need the equivalent of 8 bits precision. An 8 bit digital multiplier needs about a thousand transistors. The other major advantage is that you don't need to move data from memory to the processor - once you put your data into memory, you can do computation right there on the spot ('in memory computing').