Nothing will happen, we already have kind of translation activities happen between "incompatible views" in real life, it works.
Besides, in weights there are no real incompatibility can happen, all is unified with same math chain through architecture. Different set of local minimums/maximums can have different strong/weak consequences but nothing incompatible.
Different consistent representations also mean that model probably can be simplier for the same task, sign of excessive capacity, imho
Besides, in weights there are no real incompatibility can happen, all is unified with same math chain through architecture. Different set of local minimums/maximums can have different strong/weak consequences but nothing incompatible.
Different consistent representations also mean that model probably can be simplier for the same task, sign of excessive capacity, imho