This is a very simple RNN built on top of Theano.

 - No learning methods.
 - No error functions.
 - Only inputs, hidden, outputs, bias for hidden.
 - Recurrent weights only from hidden to hidden.
 - Only tanh, sig and identity transfer function.
 - Long short-term memory cells.

Will be extended, depending on what Theano offers (they plan to add BPTT) and
how much time I have.

Performance: Quick tests show that this implementation is roughly 20% as fast as
PyBrain with arac.

Usage:

    >>> import scipy
    >>> from rnn import RecurrentNetwork

Create a network with three hiddens. The transfer functions can be specified as
'sig', 'tanh' and 'id'.

    >>> net = RecurrentNetwork(1, 3, 1, hiddenfunc='tanh', outfunc='id')

If you want to use LSTM cells, call

    >>> net = LstmNetwork(1, 3, 1, outfunc='id')

Generate a sequence of length 3.

    >>> inpt = scipy.random.random((3, 1))

Activate sequence in the network.

    >>> net(inpt)
    [array([[-0.86715716, -0.38825977, -0.8660872 ],
           [-0.15247899,  0.787628  , -0.88322586],
           [-0.95365858, -0.55350506, -0.89969933]], dtype=float32), array([[ 1.33543777],
           [-0.95650125],
           [ 1.68285382]], dtype=float32)]

First list item is the hidden activations, second is the results. In the case of
LSTMs, there is three arrays (state, hidden, output).