Google patents method for designing image processing layers

1. A computer-implemented method, comprising: inputting an input comprising image data into a machine-learned model, the machine-learned model comprising a backbone model and a pyramid layer; processing the input using the backbone model to generate an initial plurality of output feature representations; processing, by the pyramid layer, a plurality of input feature representations based on the initial plurality of output feature representations to generate one or more output feature representations; wherein the pyramid layer combines a first feature representation and a second feature representation to generate a third feature representation, wherein: the first feature representation corresponds to a first resolution; the second feature representation corresponds to a second resolution different from the first resolution; and the third feature representation corresponds to a third resolution different from the first resolution and the second resolution; and generating, based on the third feature representation, an output from the machine-learned model. A computing system comprising: one or more processors; a controller model configured to generate new architectures for a pyramid layer that receives a plurality of input feature representations output by a backbone model and, in response, outputs a plurality of output feature representations, wherein the plurality of input feature representations have a plurality of different input resolutions and wherein the plurality of output feature representations have a plurality of different output resolutions; and one or more non-transitory computer-readable media that collectively store instructions that, when executed by the one or more processors, cause the computing system to perform operations, the operations comprising: for each of a plurality of iterations: receiving a new pyramid layer architecture as an output of the controller model; and evaluating one or more performance characteristics of a machine-learned pyramidal feature model that comprises the backbone model and one or more pyramid layers that have the new pyramid layer architecture.