New approach to the storage capacity of neural networks using the minimum distance between input patterns

This paper presents the new derivation of the storage capacity of neural networks with binary weights w_i∈{0,1},{-1,+1}. Our approaches are based on introducing a new parameter “d” (minimum distance between input patterns), not a usual parameter “p” (number of input patterns). Taking a new parameter “d” to characterize the input patterns, some results on the information theory can be applied to the computation of the storage capacity of neural networks with binary weights. This approach succeed to obtain almost the same storage capacities as those by the replica method in statistical physics