That’s part of the idea with the whole mixture of experts (MoE) approach in newer models actually.
Rather than using a single neural net that’s say 512 wide, you split it into eight channels/experts of 64. If the neural net can pick the correct channel for each inference, then you only have to run 1/8th of the neurons on every forward pass. Of course, once you have your 8 channels/experts in parallel, you now need to decide which expert/channel to use for each token you want to process. This is called a router which takes in an input and decides which expert/channel to send it to. The router itself is a tiny neural network. It is a matrix that converts the input vectors to a router choice. And the router itself has a small set of trainable weights that gets trained together with the MoE.
That’s part of the idea with the whole mixture of experts (MoE) approach in newer models actually.
Rather than using a single neural net that’s say 512 wide, you split it into eight channels/experts of 64. If the neural net can pick the correct channel for each inference, then you only have to run 1/8th of the neurons on every forward pass. Of course, once you have your 8 channels/experts in parallel, you now need to decide which expert/channel to use for each token you want to process. This is called a router which takes in an input and decides which expert/channel to send it to. The router itself is a tiny neural network. It is a matrix that converts the input vectors to a router choice. And the router itself has a small set of trainable weights that gets trained together with the MoE.