Deep neural networks are a compelling attempt at recreating the underlying mechanisms of the human brain. Though we seem to be on the road to artificial general intelligence, several obstacles remain.
The most interesting hurdle to me is the limited functional ability of DNNs. For example, a state-of-the-art facial identification network of great neuronal capacity can only answer one binary question: are these two faces the same person?
To become generally intelligent, we need to increase the functional capacity of networks. Networks need to be able to serve many disparate functions, learning to cluster neurons for each function as well as a meta-cluster for selection.
A network could eventually be allowed to learn its functions entirely and finally, with active inference, direct its own fate.