MPEG-4 AAC uses prediction and Long Term Prediction. What are their differences?
And, when are they used?

LTP is conducted in the time domain audio samples whereas backward prediction is conducted in the spectral domain.

thx much...
are there any more differences? what are the algorithms for those?

Well , some parts of the spectrals are pretty stationary.. For example, a tone signal would remain very much the same across a few audio frames. As a result we could reconstruct these spectrals based on previous frames. So we don't need to code them again in the present frame! This would improve the coding efficiency of the overall algorithm. This is the Backward prediction.

The Long Term Perdiction on the other hand removes periodicy in the time domain signal. Some audio signal especially human speech has something called the "pitch information". LTP would be able to efficiently preserved this pitch information so that the coded audio would not sound "flat or dull" especially at very low bitrates.

so it means we use the LTP in speech usually?

and i read that the computation complexity of LTP is lower than backward prediction, is that correct? how do they compare?

Yes, speech especially. That is what is noted in the MPEG4 specs.
I haven't really find out personally about the complexity issue.

Also LTP can work on both long block and short block unlike backward prediction which is switched off when changed to short block.