2 ms·
I think the missing part - symbolic reasoning - can be handled by graph neural nets and attention mechanisms. Graph neural nets take as input a set of objects a
by visarga 8y ago
I think the missing part - symbolic reasoning - can be handled by graph neural nets and attention mechanisms. Graph neural nets take as input a set of objects and their relations and predict information about the objects, their relations and the whole graph. They learn invariance in relation space, while being permutation invariant in the object space. This kind of invariance is what we need to improve on. We already have some promising attempts such as the GCN. This and creating environments for artificial agents - a child AI needs a place for playing and experimentation, not a mere static dataset.