Google patents neural networks that repeat steps until results stabilize

1. A method performed by one or more data processing apparatuses for performing a machine learning task using a neural network that comprises a sequence of blocks, wherein the sequence of blocks comprises one or more first blocks followed by one or more intermediate blocks followed by one or more last blocks, and wherein the method comprises: receiving a network input for a machine learning task; processing the network input through the sequence of blocks to generate a network output for the machine learning task, the processing comprising: for the one or more first blocks, processing the network input using the one or more first blocks to generate a first block output; for each of the one or more intermediate blocks, processing an intermediate block input to generate a final intermediate block output, comprising repeatedly performing looping iterations until a loop termination criterion is satisfied, each looping iteration comprising: obtaining a current intermediate block input for the looping iteration, wherein the current intermediate block input for a first looping iteration is the intermediate block input, and wherein the current intermediate block input for each subsequent looping iteration is a current intermediate block output generated by the intermediate block for a preceding looping iteration, and processing the current intermediate block input for the looping iteration using at least the intermediate block to generate a current intermediate block output for the looping iteration; and for the one or more last blocks, processing the final intermediate block output generated by a last intermediate block using the one or more last blocks to generate a network output for the machine learning task.