New for: D1, D2, D3, D4, D5
dynamical systems and have recently become widely used in the "deep
learning" field of machine learning, especially for speech and language
processing tasks. For instance, Google's speech recognition and language
translation services are based on RNNs.
However, the deep learning set-ups for RNN training are computationally
expensive, require very large volumes of training data, and need
high-precision numerical processing. For such reasons, deep-learning
variants of RNNs are problematic in fields where training data are
scarce, where fast and cheap algorithms are desired, or where noisy or
low-precision hardware is to be used. This is often the case in domains
of nonlinear signal processing, control, brain-machine interfacing, or
biomedical signal processing.
Reservoir Computing (RC) is an alternative machine learning approach for
RNNs which is in many aspects complementary to the ways of deep
learning. In RC, a large, random, possibly low-precision and noisy RNN
is used as a nonlinear excitable medium - called the "reservoir" - which
is driven by an input signal. The reservoir itself is not adapted or
trained. Instead, only a "readout" mechanism is trained, which assembles
the desired output signal from the large variety of random, excited
signals within the reservoir. This readout training is cheap - typically
just a linear regression. RC has become a popular approach in research
that aims at useful computations on the basis on unconventional hardware
(non-digital, noisy, low-precision).
The talk gives a quick introduction to the basic principles of RC, and
then proceeds to recent add-ons to that paradigm.