Iterative neural networks for adaptive inference on resource-constrained devices

The computational cost of evaluating a neural network usually only depends on design choices such as the number of layers or the number of units in each layer and not on the actual input. In this work, we build upon deep Residual Networks (ResNets) and use their properties to design a more efficient adaptive neural network building block. We propose a new architecture, which replaces the sequential layers with an iterative structure where weights are reused multiple times for a single input image, reducing the storage requirements drastically. In addition, we incorporate an adaptive computation module that allows the network to adjust its computational cost at run time for each input sample independently. We experimentally validate our models on image classification, object detection and semantic segmentation tasks and show that our models only use their full capacity for the hardest input samples and are more efficient on average.