The DEtection TRansformer (DETR) approach, which uses a transformer encoder-decoder architecture and a set-based global loss, has become a building block in many transformer based applications. However, as originally presented, the assignment cost and the global loss are not aligned, i.e., reducing the former is likely but not guaranteed to reduce the latter. And the issue of gradient is ignored when a combinatorial solver such as Hungarian is used. In this paper we show that the global loss can be expressed as the sum of an assignment-independent term, and an assignment-dependent term which can be used to define the assignment cost matrix. Recent results on generalized gradients of optimal assignment cost with respect to parameters of an assignment problem are then used to define generalized gradients of the loss with respect to network parameters, and backpropagation is carried out properly. Our experiments using the same loss weights show interesting convergence properties and a potential for further performance improvements.

When samples have internal structure, we often see a mismatch between the objective optimized during training and the model's goal during inference. For example, in sequence-to-sequence modeling we are interested in high-quality translated sentences, but training typically uses maximum likelihood at the word level. Learning to recognize individual faces from group photos, each captioned with the correct but unordered list of people in it, is another example where a mismatch between training and inference objectives occurs. In both cases, the natural training-time loss would involve a combinatorial problem -- dynamic programming-based global sequence alignment and weighted bipartite graph matching, respectively -- but solutions to combinatorial problems are not differentiable with respect to their input parameters, so surrogate, differentiable losses are used instead. Here, we show how to perform gradient descent over combinatorial optimization algorithms that involve continuous parameters, for example edge weights, and can be efficiently expressed as integer, linear, or mixed-integer linear programs. We demonstrate usefulness of gradient descent over combinatorial optimization in sequence-to-sequence modeling using differentiable encoder-decoder architecture with softmax or Gumbel-softmax, and in weakly supervised learning involving a convolutional, residual feed-forward network for image classification.

Learning in Riemannian orbifolds is motivated by existing machine learning algorithms that directly operate on finite combinatorial structures such as point patterns, trees, and graphs. These methods, however, lack statistical justification. This contribution derives consistency results for learning problems in structured domains and thereby generalizes learning in vector spaces and manifolds.