Slide 11 of 22

Notes:

• From that point onwards, instead of basing the prediction on the decoded samples of the same iteration, we base the prediction on decoded samples of the previous iteration. This way, if in the current iteration, we badly quantize a frame, it will have no effect on any of the consecutive frames since their prediction is based on decoded samples of the previous iteration and not this iteration. We generate new training residuals on which we design a new quantizer and reiterate this process.