Reservoir computing: Adaptive online machine learning and neural networks

Neri van Otten, Data Scientist at Conversocial at Data Science London @ds_ldn meetup March 28th, 20123

Transcript

1. Neri Van Otten neri.vvo@gmail.com
2. What is Reservoir Computing?  Temporal Recurrent Neural Network (Dominey 1995)  Liquid State Machines (Natschläger, Maass and Markram 2002)  Echo State Networks (Jaeger 2001)  Decorrelation-Backpropagation Learning (Steil 2004)
3. Types of Neural Networks
4. Recurrent Neural Networks (RNN)  RNN attractive solution to many engineering problems  learn by example  model highly nonlinear systems  Training phase is slow and unstable  Reservoir Computing avoids this problem by not training it at all
5. A Reservoir is like Water
6. Reservoir Computing (RC)
7. RC Explained  Large random RNN as an excitable medium  Driven by input signals, each unit in the RNN creates its own nonlinear transform of the input  Output signals are read out from excited RNN  Typically a simple linear combination of the reservoir signals  Outputs trained in supervised way  Typically by linear regression of the teacher output on the tapped reservoir signals.
8. Advantages of RC  Training phase is computationally fast  Works as a dynamic system just as other RNN  Form of memory is used  Suited for a wide variety of problems:  epileptic seizure detection  brain computing interfaces  time series prediction
9. Feedback Controller  “Feedback control by online learning an inverse model” (Waegeman 2012)
10. Training
11. Testing
12. Feedback Controller
13. Use RC with OGER  > import Oger  > resnode = Oger.nodes.ReservoirNode(output_dim = 100)  > readoutnode = Oger.nodes.RidgeRegressionNode()  > flow = resnode + readoutnode  > flow.train(data)