5.
J. Schmidhuber, D. Wierstra, M. Gagliolo, F. Gomez.
Training Recurrent Networks by Evolino.
Neural Computation, 19(3): 757779, 2007.
PDF.
4.
H. Mayer, F. Gomez, D. Wierstra, I. Nagy, A. Knoll, and J. Schmidhuber (2006).
A System for Robotic Heart Surgery that Learns to Tie Knots Using Recurrent
Neural Networks. Proc. IROS06, Beijing.
PDF.
3.
J. Schmidhuber, D. Wierstra, F. J. Gomez.
Evolino: Hybrid Neuroevolution / Optimal Linear Search
for Sequence Learning.
Proceedings of the 19th International Joint Conference
on Artificial Intelligence (IJCAI), Edinburgh, p. 853858, 2005.
PDF.
2.
D. Wierstra, F. Gomez, J. Schmidhuber.
Modeling systems with internal state using Evolino.
In Proc. of the 2005 conference on genetic and
evolutionary computation (GECCO), Washington, D. C.,
pp. 17951802, ACM Press, New York, NY, USA, 2005.
PDF.
Got a GECCO best paper award.
1.
J. Schmidhuber, M. Gagliolo, D. Wierstra, F. Gomez.
Evolino for Recurrent Support Vector Machines.
TR IDSIA1905, v2, 15 Dec 2005.
PDF.
Short version at ESANN 2006.



Basic principle: Evolve
an RNN population; to obtain some RNN's fitness DO:
Feed the training sequences into the RNN. This yields sequences of hidden unit
activations. Compute an optimal linear mapping from hidden to target trajectories.
The fitness of the recurrent hidden units is the RNN
performance on a validation set, given this mapping.
If the goal is to minimize mean squared error, then
use the pseudoinverse for computing the optimal mapping (left).
If the goal is to maximize the margin, then
use quadratic programming. This yields
Recurrent Support Vector Machines.
A recent journal publication on an
EVOLINO application to Robotics:
H. Mayer, F. Gomez, D. Wierstra, I. Nagy, A. Knoll, and J. Schmidhuber.
A System for Robotic Heart Surgery that Learns to Tie Knots Using Recurrent Neural Networks.
Advanced Robotics,
22/1314, p. 15211537, 2008, in press.



Experiments:
EVOLINO trained
LSTM RNNs
with coevolving hidden neurons
can learn to predict several time series that Echo State nets (Jaeger,
Science 2004)
and other traditional RNNs
cannot learn, such as superimposed sines out of phase (right)
or certain input streams based on grammatical rules.
Fibonacci web design
by J. Schmidhuber


EVOLINObased LSTM was able to learn up to 5 sines (above: 3 sines, with
zoom on detail),
certain contextsensitive grammars, and
the Mackey Glass time series (top of this page), which is not
a very good RNN benchmark though, since even feedforward
nets can learn it well. Sometimes EvolinoLSTM outperformed
GradientLSTM on problems where
LSTMbased systems were the only competitors (since
traditional RNNs or HMMs did not work).
 
