A Constrained Optimization Approach to Preserving Prior Knowledge During Incremental Training