Random projections and matrix pseudoinversion for neural training