GENERATING DIFFERENTIABLE ORDER STATISTICS USING SORTING NETWORKS

1 . A method performed by one or more data processing apparatus for generating one or more differentiable order statistics for a vector of scores, the method comprising:

obtaining the vector of scores, wherein each position in the vector of scores is associated with a respective index from a set of indices;

obtaining a plurality of pairs of indices, wherein each pair of indices comprises a respective first index and a respective second index from the set of indices;

generating a respective swapping probability for each pair of indices based on the vector of scores;

generating, for each pair of indices, a respective soft-swapping matrix for the pair of indices as a combination of: (i) an identity matrix that, if applied to the vector of scores, would leave the vector of scores unchanged, and (ii) an exchange matrix that, if applied to the vector of scores, would swap a pair of scores indexed by the pair of indices, wherein the exchange matrix is weighted in the combination by the swapping probability for the pair of indices; and

generating the one or more differentiable order statistics for the vector of scores using the soft-swapping matrices.

2 . The method of claim 1 , wherein obtaining the vector of scores comprises generating each score in the vector of scores as an output of a machine learning model having a plurality of machine learning model parameters.

3 . The method of claim 2 , further comprising training the machine learning model parameters using an objective function that depends on the differentiable order statistics, comprising:

determining gradients of the differentiable order statistics with respect to the machine learning model parameters; and

updating values of the machine learning model parameters using the gradients of the differentiable order statistics with respect to the machine learning model parameters.

4 . The method of claim 3 , wherein the differentiable order statistics include a top-K order statistic that defines whether a target index is included in a set of indices of the K highest scores from the vector of scores.

5 . The method of claim 4 , wherein the machine learning model comprises a neural network.

6 . The method of claim , wherein the respective swapping probability for each pair of indices is strictly greater than zero and strictly less than one.

7 . The method of claim 1 , wherein for each pair of indices, the soft-swapping matrix for the pair of indices is generated as:

(1−ρ)I+ρE,

wherein ρ is the swapping probability, I is the identity matrix, and E is the exchange matrix.

8 . The method of claim 1 , wherein the plurality of pairs of indices are associated with a sequential ordering, wherein the swapping probabilities for the pairs of indices are sequentially generated starting from a first pair of indices in the sequential ordering of the pairs of indices, and wherein for each pair of indices, generating the swapping probability for the pair of indices comprises:

determining the swapping probability for the pair of indices based on a pair of scores indexed by the pair of indices in a current copy of the vector of scores; and

applying a swapping operator associated with the pair of indices to the current copy of the vector of scores, wherein the swapping operator is from a sequence of swapping operators that define a sorting network.

9 . The method of claim 8 , wherein applying the swapping operator associated with the pair of indices to the current copy of the vector of scores comprises:

evaluating an inequality between the pair of scores indexed by the pair of indices in the current copy of the vector of scores; and

in response to determining that the inequality is satisfied, swapping the pair of scores indexed by the pair of indices in the current copy of the vector of scores.

10 . The method of claim 8 , wherein generating the swapping probability for a final pair of indices comprises applying a final swapping operator to the current copy of the vector of scores, and wherein after applying the final swapping operator, the scores in the current copy of the vector of scores are sorted.

11 . The method of claim 8 , wherein for each pair of indices, determining the swapping probability for the pair of indices based on the pair of scores indexed by the pair of indices in a current copy of the vector of scores, comprises:

determining a difference between the pair of scores indexed by the pair of indices in the current copy of the vector of scores;

determining a product of an inverse of a dispersion factor and the difference between the pair of scores indexed by the pair of indices in the current copy of the vector of scores; and

determining the swapping probability by applying a sigmoid function to a result of the product.

12 . The method of claim 8 , further comprising obtaining a sorted version of the vector of scores; and

wherein for each pair of indices, determining the swapping probability for the pair of indices based on the pair of scores indexed by the pair of indices in the current copy of the vector of scores, comprises:

determining an error between: (i) the pair of scores indexed by the pair of indices in the current copy of the vector of scores, and (ii) a pair of scores indexed by the pair of indices in the sorted version of the vector of scores; and

determining the swapping probability based on the error.

13 . The method of claim 12 , wherein for each pair of indices, determining the error between: (i) the pair of scores indexed by the pair of indices in the current copy of the vector of scores, and (ii) the pair of scores indexed by the pair of indices in the sorted version of the vector of scores, comprises:

determining a cost matrix [C ij ] i,j=1,2 , where C ij =h(y i −x j ), [x 1 , x 2 ] are the pair of scores indexed by the pair of indices in the current copy of the vector of scores, [y 1 , y 2 ] are the pair of scores indexed by the pair of indices in the sorted version of the vector of scores, and h(⋅) is a convex and non-negative function.

14 . The method of claim 13 , wherein for each pair of indices, determining the swapping probability based on the error comprises determining the swapping probability λ as:

λ

=

exp

⁡

(

C

i

⁢

i

+

C

jj

ϵ

)

exp

⁡

(

C

i

⁢

i

+

C

jj

ϵ

)

+

exp

⁡

(

C

ij

+

C

j

⁢

i

ϵ

)

where ϵ is a dispersion factor.

15 . The method of claim 1 , wherein generating the one or more differentiable order statistics for the vector of scores using the soft-swapping matrices comprises:

generating a matrix product of the soft-swapping matrices; and

generating the one or more differentiable order statistics based on the matrix product of the soft-swapping matrices.

16 . The method of claim 15 , wherein the one or more differentiable order statistics comprise a sorted version of the vector of scores, and wherein generating the sorted version of the vector of scores comprises:

determining a product of: (i) the matrix product of the soft-swapping matrices, and (ii) the vector of scores.

17 . The method of claim 15 , wherein the one or more differentiable order statistics comprise a maximum score in the vector of scores, and wherein generating the maximum score in the vector of scores comprises:

determining a product of: (i) a transpose of a final column of an identity matrix, (ii) the matrix product of the soft-swapping matrices, and (iii) the vector of scores.

18 . The method of claim 15 , wherein the one or more differentiable order statistics comprise a minimum score in the vector of scores, and wherein generating the minimum sore in the vector of scores comprises:

determining a product of: (i) a transpose of a first column of an identity matrix, (ii) the matrix product of the soft-swapping matrices, and (iii) the vector of scores.

19 . One or more non-transitory computer storage media storing instructions that when executed by one or more computers cause the one or more computers to perform operations for generating one or more differentiable order statistics for a vector of scores, the operations comprising:

obtaining the vector of scores, wherein each position in the vector of scores is associated with a respective index from a set of indices;

obtaining a plurality of pairs of indices, wherein each pair of indices comprises a respective first index and a respective second index from the set of indices;

generating a respective swapping probability for each pair of indices based on the vector of scores;

generating, for each pair of indices, a respective soft-swapping matrix for the pair of indices as a combination of: (i) an identity matrix that, if applied to the vector of scores, would leave the vector of scores unchanged, and (ii) an exchange matrix that, if applied to the vector of scores, would swap a pair of scores indexed by the pair of indices, wherein the exchange matrix is weighted in the combination by the swapping probability for the pair of indices; and

generating the one or more differentiable order statistics for the vector of scores using the soft-swapping matrices.

20 . A system comprising:

one or more computers; and

one or more storage devices communicatively coupled to the one or more computers, wherein the one or more storage devices store instructions that, when executed by the one or more computers, cause the one or more computers to perform operations for generating one or more differentiable order statistics for a vector of scores, the operations comprising:

obtaining the vector of scores, wherein each position in the vector of scores is associated with a respective index from a set of indices;

obtaining a plurality of pairs of indices, wherein each pair of indices comprises a respective first index and a respective second index from the set of indices;

generating a respective swapping probability for each pair of indices based on the vector of scores;

generating, for each pair of indices, a respective soft-swapping matrix for the pair of indices as a combination of: (i) an identity matrix that, if applied to the vector of scores, would leave the vector of scores unchanged, and (ii) an exchange matrix that, if applied to the vector of scores, would swap a pair of scores indexed by the pair of indices, wherein the exchange matrix is weighted in the combination by the swapping probability for the pair of indices; and

generating the one or more differentiable order statistics for the vector of scores using the soft-swapping matrices.