A connection represents the synaptic connection between Units. It contains a synaptic weight value that determines how much one unit affects the other. The adjustment of this weight by a Learning Algorithm is how the network learns.
Individual connections between units are controlled by an overall Projection that determines the overall pattern of connectivity.
As of Version 6.4.0, connection-level data is no longer stored using Connection objects -- instead the members of the connection are allocated in contiguous arrays (i.e., all weights, then all dwts, etc) to optimize performance. See Unit for details on how to access connections in Program code.