We propose to investigate the feasibility of "computational reduction" of large-scale neural networks, aiming at providing with fine control over the trade-off between the number of required floating point operations-and correspondingly the required hardware capabilities-and the statistical accuracy of the prediction.
