Simply pruning connections with near zero weights doesn't help with the original matrix multiplication approach to CNNs. We will need to use optimized sparse matrix multiplication in order to see benefits from pruning. Will need to deal with the tradeoff between the benefits of pruning and overhead of dealing with sparse matrices.
Abandon
This is a common approach (thinking) which is widely adopted in the distributed systems to reduces the overhead of the networks. The parameter server paper from Mu Li et.al in OSDI'14 uses similar ideas.
Simply pruning connections with near zero weights doesn't help with the original matrix multiplication approach to CNNs. We will need to use optimized sparse matrix multiplication in order to see benefits from pruning. Will need to deal with the tradeoff between the benefits of pruning and overhead of dealing with sparse matrices.
This is a common approach (thinking) which is widely adopted in the distributed systems to reduces the overhead of the networks. The parameter server paper from Mu Li et.al in OSDI'14 uses similar ideas.