News for hackers. Stuff that matters.
Re:Predictive Coding Approximates Backprop Along Arbitrary C... by nl on Tuesday October 06, 2020 @07:40AM (#24696676)
This maybe true, but on existing hardware isn't super useful.

Even on the potential neuromophic hardware it's "only" parallel with the number of layers. That's useful but not as parallel as current approaches.

Re:Predictive Coding Approximates Backprop Along Arbitrary C... by p1esk on Wednesday October 07, 2020 @01:14PM (#24710389)
What do you mean? Looking at the equations, I don't see why this local learning rule can't be parallel with the number of neurons, so it's vastly more parallel, but requires multiple iterations per neuron (100x in the paper).