Anytime Neural Network: a Versatile Trade-off Between Computation and Accuracy

Anytime predictors first produce crude results quickly, and then continuously refine them until the test-time computational budget is depleted. Such predictors are used in real-time vision systems and streaming-data processing to efficiently utilize varying test-time budgets, and to reduce average prediction cost via early-exits. However, anytime prediction algorithms have difficulties utilizing the accurate predictions of deep neural networks (DNNs), because DNNs are often computationally expensive without competitive intermediate results. In this work, we propose to add auxiliary predictions in DNNs to generate anytime predictions, and optimize these predictions simultaneously by minimizing a carefully constructed weighted sum of losses, where the weights also oscillate during training. The proposed anytime neural networks (ANNs) produce reasonable anytime predictions without sacrificing the final performance or incurring noticeable extra computation. This enables us to assemble a sequence of exponentially deepening ANNs, and it achieves, both theoretically and practically, near-optimal anytime predictions at every budget after spending a constant fraction of extra cost. The proposed methods are shown to produce anytime predictions at the state-of-the-art level on visual recognition data-sets, including ILSVRC2012.