Soumen Madan

One DNN that can recognize images and other DNN that can recognize sounds, add them in a package and you would get a package having different models. Won't this be an example of modularity?


what about adding a "spacial" dimension to the neurons in the the network and a dying a cost for the distance axions have to be formed for connections.
could this lead to higher modularity with better bunching of the neurons related to one other for the same task. as neurons in brains tend to bunch in 100 to 100.000 neurons with multiple layers in the column, with some neurons expressed more vertically (they feed back in the same column on multiple levels even back closer to the signal start), or horizontally feeding signals to close column (or clusters) on the same level, or the most abundant in to the same level of the same cluster feeding in to itself.