Scale-permuted machine learning architecture
A computer-implemented method of generating scale-permuted models can generate models having improved accuracy and reduced evaluation computational requirements. The method can include defining, by a computing system including one or more computing devices, a search space including a plurality of candidate permutations of a plurality of candidate feature blocks, each of the plurality of candidate feature blocks having a respective scale. The method can include performing, by the computing system, a plurality of search iterations by a search algorithm to select a scale-permuted model from the search space, the scale-permuted model based at least in part on a candidate permutation of the plurality of candidate permutations.
1. A computer-implemented method of generating scale-permuted models having improved accuracy or reduced computational requirements, the method comprising:
defining, by a computing system comprising one or more computing devices, a search space including a plurality of candidate permutations of a plurality of candidate feature blocks, each of the plurality of candidate feature blocks having a respective resolution;
performing, by the computing system, a plurality of search iterations by a search algorithm to select a scale-permuted model from the search space,
wherein the scale-permuted model comprises a sequence of blocks comprising:
a first feature block in the sequence having a first resolution,
a second feature block next in the sequence after the first feature block, the second feature block having a second resolution higher than the first resolution, and
a third feature block next in the sequence after the second feature block, the third feature block having a third resolution lower than the second resolution and different from the first resolution, and
wherein the scale-permuted model is based at least in part on a candidate permutation of the plurality of candidate permutations, the candidate permutation comprising a plurality of permuted feature blocks having a permuted ordering that differs from an initial ordering of the plurality of candidate feature blocks; and
providing, by the computing system, the scale-permuted model as an output.
2. The computer-implemented method of claim 1 , wherein performing each of the plurality of search iterations comprises:
determining, by the computing system, a candidate scale-permuted model from the search space, the candidate scale-permuted model comprising the plurality of permuted feature blocks based at least in part on the candidate permutation; and
evaluating, by the computing system, the candidate scale-permuted model based at least in part on a performance estimation strategy to obtain an evaluation of the candidate scale-permuted model;
wherein the scale-permuted model is selected based at least in part on the evaluations of the candidate scale-permuted model for each of the plurality of search iterations.
3. The computer-implemented method of claim 2 , wherein determining the candidate scale-permuted model from the search space comprises:
determining, by the computing system, the plurality of permuted feature blocks based at least in part on the candidate permutation; and
determining, by the computing system, one or more cross-block connections between the plurality of permuted feature blocks.
4. The computer-implemented method of claim 3 , wherein the one or more cross-block connections comprises at least one cross-scale connection configured to connect a parent block of the plurality of permuted feature blocks, the parent block having a first resolution, to a target block of the plurality of permuted feature blocks, the target block having a second resolution.
5. The computer-implemented method of claim 2 , wherein determining the candidate scale-permuted model from the search space comprises applying, by the computing system, one or more block adjustments to the plurality of permuted feature blocks.
6. The computer-implemented method of claim 5 , wherein the one or more block adjustments comprise at least one type adjustment.
7. The computer-implemented method of claim 5 , wherein the one or more block adjustments comprise at least one resolution adjustment.
8. The computer-implemented method of claim 1 , wherein the plurality of candidate feature blocks is defined based at least in part on an existing model architecture.
9. The computer-implemented method of claim 1 , wherein the scale-permuted model comprises a fourth feature block ordered subsequent to the third feature block, the respective resolution of the fourth feature block being higher than the respective resolution of the third feature block.
10. A computer-implemented method of generating scale-permuted models having improved accuracy and reduced evaluation computational requirements, the computer-implemented method comprising:
receiving, at a computing system comprising one or more computing devices, a plurality of candidate feature blocks from a user, each of the plurality of candidate feature blocks having a respective resolution;
defining, by the computing system, a search space including a plurality of candidate permutations of the plurality of candidate feature blocks;
performing, by the computing system, a plurality of search iterations by a search algorithm to select a scale-permuted model from the search space,
wherein the scale-permuted model comprises a sequence of blocks comprising:
a first feature block in the sequence having a first resolution,
a second feature block next in the sequence after the first feature block, the second feature block having a second resolution higher than the first resolution, and
a third feature block next in the sequence after the second feature block, the third feature block having a third resolution lower than the second resolution and different from the first resolution;
wherein the scale-permuted model is based at least in part on a candidate permutation of the plurality of candidate permutations;
wherein performing each of the plurality of search iterations comprises:
determining, by the computing system, a candidate scale-permuted model from the search space, the candidate scale-permuted model comprising a plurality of permuted feature blocks based at least in part on the candidate permutation, the plurality of permuted feature blocks having a permuted ordering that differs from an initial ordering of the plurality of candidate feature blocks; and
evaluating, by the computing system, the candidate scale-permuted model based at least in part on a performance estimation strategy to obtain an evaluation of the candidate scale-permuted model; and
wherein the scale-permuted model is selected based at least in part on the evaluations of the candidate scale-permuted model for each of the plurality of search iterations; and
providing, by the computing system, the scale-permuted model to the user.
11. The computer-implemented method of claim 10 , wherein determining the candidate scale-permuted model from the search space comprises:
determining, by the computing system, the plurality of permuted feature blocks based at least in part on the candidate permutation; and
determining, by the computing system, one or more cross-block connections between the plurality of permuted feature blocks.
12. The computer-implemented method of claim 10 , wherein determining the candidate scale-permuted model from the search space comprises applying, by the computing system, one or more block adjustments to the plurality of permuted feature blocks.
13. The computer-implemented method of claim 10 , wherein the search algorithm comprises Neural Architecture Search.