4 ms·
You might be interested in the SNARC [1], a very early neural network machine. Each neuron was a few vacuum tubes, weights were stored in potentiometers, and tr
by wiml 5y ago
You might be interested in the SNARC [1], a very early neural network machine. Each neuron was a few vacuum tubes, weights were stored in potentiometers, and training was done through a system of belts and electromechanical clutches to adjust potentiometers based on its response to training data.
[1] http://cyberneticzoo.com/mazesolvers/1951-maze-solver-minsky-edmonds-american/ http://cyberneticzoo.com/mazesolvers/1951-maze-solver-minsky...
- throwaway316943 5y agoYes this is the type of implementation that inspired the idea. Now with cheap and tiny microcontrollers each neuron can be self contained and handle it’s own weights without need for external memory or components. The GreenArray chip mentioned by another poster seems to fit this though each core is only directly connected to 4 others you could work around that limitation by having a layer of routing cores that pass messages from one neuron layer to the next.