Multi-Scale Dense Convolutional Networks for Efficient Prediction

We introduce a new convolutional neural network architecture with the ability to adapt dynamically to computational resource limits at test time. Our network architecture uses progressively growing multi-scale convolutions and dense connectivity, which allows for the training of multiple classifiers at intermediate layers of the network. We evaluate our approach in two settings: (1) anytime classification, where the network's prediction for a test example is progressively updated, facilitating the output of a prediction at any time; and (2) budgeted batch classification, where a fixed amount of computation is available to classify a set of examples that can be spent unevenly across "easier" and "harder" inputs. Experiments on three image-classification datasets demonstrate that our proposed framework substantially improves the state-of-the-art in both settings.