http://arxiv.org/abs/1209.3129
To our knowledge, the system presented here is the first analog readout for an experimental reservoir computer. While the results presented here are prelim- inary, and there is much optimization of experimental parameters to be done,the system already outperforms non-reservoir methods. We expect to extend easily this approach to different tasks, already studied in [9, 10], including a spoken digit recognition task on a standard dataset.3
Further performance improvements can reasonably be expected from fine- tuning of the training parameters: for instance the amount of regularization in the ridge regression procedure, that here is left constant at 1 · 10−4 , should be tuned for best performance. Adaptive training algorithms, such as the ones mentioned in [21], could also take into account nonidealities in the readout components
Moreover the choice of τ, as Figure 3 shows, is not obvious and a
more extensive investigation could lead to better performance.
The architecture proposed here is simple and quite straightforward to re-alize. It is very modular, meaning that it can be added at the output of any preexisting time multiplexing reservoir with minimal effort, whether it is based on optics or electronics. The capacitor at the end of the circuit could probably
be substituted with a more complicated, active electronic circuit performing the summation of the incoming signal before resetting itself. This would eliminate the problem of residual voltages, and allow better performance at the cost of increased complexity of the readout. The main interest of the analog readout is that it allows optoelectronic reser- voir computers to fully leverage their main characteristic, which is the speed of operation.
Indeed, removing the need for slow, offline postprocessing is indi- cated in [13] as one of the major challenges in the field. Once the training is finished, optoelectronic reservoirs can process millions of nonlinear nodes per second [10]; however, in the case of a digital readout, the node states must be recovered and postprocessed to obtain the reservoir outputs. It takes around 1.6 seconds for the digital readout in our setup to retrieve and digitize the states generated by a 9000 symbol input sequence. The analog readout removes the need for postprocessing, and can work at a rate of about 8.5 μs per input sym- bol, five orders of magnitude faster than the electronic reservoir reported in [8].
Finally, having an analog readout opens the possibility of feedback - using the output of the reservoir as input or part of an input for the successive time steps. This opens the way for different tasks to be performed [15] or different training techniques to be employed [14].
The supplementary materials pdf from the nature link http://www.nature.com/srep/
http://en.wikipedia.org/wiki/Electro-optic_modulator
Compared to the million of steps in semiconductor fabrication it's positively easy to construct one of these in relation. This paper only came out last week so obviously nobody is gonna manufacture them yet but if it beats the cost over time it could mean incredibly economical supercomputer production. These things are so simple you could 3d print them. (They've already 3d printed fiber and lasers and lenses etc. pretty easy, Fab Lab and other semiconductor projects are stuck in the mud in comparison to the ease of this process.)
