My current image from a MoE is that the base/core might be the more generic thing and that things like python are part of one expert though.
replyWith MOE you train a router designed to select which parts to activate. The router itself is a trained neural network and the 'experts' are usually not really things like 'python'. They're just the functional subspaces I described above.
replyAgain, those subspaces are all somehow inextricably correlated and live in complex superposition spread throughout the manifold. The router doesn't know (or care) WHY those sections get lit up it just learns which ones to activate to optimize it's own reward function. So maybe it learns to activate "logic", "python" and "cupcake recipes in esperanto" whenever it see's something that kind of looks like python. It's not the best answer, it's just the best answer the tiny router could figure out.
It's all wildly complicated and inefficient, and works nothing like any reasonable human would imagine that it SHOULD operate.
I mean the only way to separate them would be to separate them at training would it not? If there is zero python in the training set, then you have a smaller training set, should be able to have fewer weights in the model
replyThere was some paper about routing at training bio-knowledge into a particular region of the model, which you then can cutoff when serving. But you probably lose some efficiency since maybe you sized that region too small/too big.
replyThat's a very clever approach. Any idea about the papers title or authors? I'd love to look it up.
reply