AutoRegNN # Luc Döbereiner AutoRegNN1 is an autoregressive feed-forward neural network that predicts the next sample based on previously generated ones. Learning happens online and in real-time while synthesizing sound. Implemented as a unit generator (UGen) for SuperCollider, it uses autoregression and backpropagation with ridge regression regularization to stabilize learning and prevent exploding gradients. The network performs two forward passes for each audio sample, using both external ...
