1. RElu, PRElu (parametrized RElu) activation functions, see https://arxiv.org/pdf/1502.01852.pdf

2. with global flag, change all BP-based algos to support cross_entropy error: if (cross_entropy==0){# cross_entropy==1 means omit f' in output layer
3. dropout dropout[2]=0.2; 	    if (rand()<dropout[rfi]) {#skip this unit for next layer
4. test (P)RElu for QBPTT, task is Protein-SS-Pred

5. RMSprop: mu=0.99 -v eta=0.001 -v eps=1e0 mu1=(1-mu)
		ldx=dx[nl,ii,ih]; #gradient
		dw[nl,ii,ih]*=mu;
		dw[nl,ii,ih]+=mu1*ldx*ldx;
		w[nl,ii,ih]+=eta*ldx/(sqrt(dw[nl,ii,ih])+eps);
		dx[nl,ii,ih]=0;