This thesis presents a learning framework that integrates predictive coding and forward mode gradient estimation to enable parallelizable local learning in neural networks – designed specifically for analog hardware implementation. Unlike backpropagation, which is optimized for digital architectures and relies on partial derivative transmission against the direction of inference, this method restructures the network into locally controlled units that compute updates in the forward direction, aligning naturally with feedback control. While forward mode gradients and predictive coding are not inherently faster than backpropagation on digital systems, the combination of the two is well-suited to analog circuits, which offer potential improvements in speed and efficiency. The novel algorithm can successfully learn nonlinear functions such as XOR and sequence prediction tasks by forecasting video frames. These results mark an important step toward building fully online, analog-learning systems capable of real-time adaptation.
Cedar Frederic Turek (Thu,) studied this question.