State Prediction: A Constructive Method to Program Recurrent Neural Networks
We introduce a novel technique to program desired state sequences into recurrent neural networks in one shot. The basic methodology and its scalability to large and input-driven networks is demonstrated by shaping attractor landscapes, transient dynamics and programming limit cycles. The approach unifies programming of transient and attractor dynamics in a generic framework.
Keywordsrecurrent neural networks input-driven dynamics learning
Unable to display preview. Download preview PDF.
- 3.Reinhart, R.F., Steil, J.J.: A Constrained Regularization Approach for Input-Driven Recurrent Neural Networks. Differ. Equ. Dyn. Syst., 1–20 (2010)Google Scholar
- 4.Reinhart, R.F., Steil, J.J.: Reservoir Regularization Stabilizes Learning of Echo State Networks with Output Feedback. In: European Symp. on ANNs (2011)Google Scholar
- 5.Jaeger, H.: Reservoir Self-Control for Achieving Invariance Against Slow Input Distortions. Technical report, Jacobs University (2010)Google Scholar
- 6.Lecun, Y., Cortes, C., http://yann.lecun.com/exdb/mnist/