For example, the article mentions the ability of clustered synapses to act independently, but , on the one hand, it has been shown independent dendrites can be approximated as an extra neural network layer (so they ARE covered by today's ANN approximation) , and OTOH there s a number of papers showing that synaptic clustering does not exist in sensory areas. And learning by rewiring is basically the introduction of random connections which persist only if their weight increases enough (roughly corresponds the continuous formation of filopodia and the fact that large spines persist longer).
Machine learning at the moment is an empirical science that has made great strides without consulting neuroscience for it. I think that has been a good thing: without having to bend towars some biological plausibility researchers have been more exploratory and creative, which has led to the creation of an empirical body of knowledge from which neuroscience could benefit in the future. OTOH, having watched the field of computational neuroscience there has not been a lot of progress since , basically the 80s. So i believe it would be best to leave each of the two fields go their own way.