Parameter learning with truncated message-passing

Training of conditional random fields often takes the form of a double-loop procedure with message-passing inference in the inner loop. This can be very expensive, as the need to solve the inner loop to high accuracy can require many message-passing iterations. This paper seeks to reduce the expense of such training, by redefining the training objective in terms of the approximate marginals obtained after message-passing is “truncated” to a fixed number of iterations. An algorithm is derived to efficiently compute the exact gradient of this objective. On a common pixel labeling benchmark, this procedure improves training speeds by an order of magnitude, and slightly improves inference accuracy if a very small number of message-passing iterations are used at test time.