Aug 8, 2017 · 1 min read
Aren’t the ideas of holographic/distributed encoding of information and information encoded as network topology inherent to connectionism?
distributed, rather than local representations on the hidden units are the natural products of connectionist training methods.
distributed representations, (unlike symbols stored in separate fixed memory locations) remain relatively well preserved when parts of the model are destroyed or overloaded
Good article. It is interesting to think of abstractions / composite representations as entanglement.
