Classifying data objects

1 . A method performed by one or more computers, the method comprising:

obtaining data comprising a respective high-dimensional representation for each category label of a plurality of category labels that each correspond to a different category of a plurality of categories, wherein the high-dimensional representation of the category label is a numeric representation of the category label in a high-dimensional space;

obtaining a new image to be classified;

processing the new image using a deep neural network to generate an aggregate high-dimensional representation of the new image from the high-dimensional representations of the plurality of category labels and respective scores for each category of one or more of the plurality of categories, wherein each of the respective scores represents a respective likelihood that the new image belongs to the corresponding category;

identifying a first term in a vocabulary of terms having a high-dimensional representation that is closest to the aggregate high-dimensional representation; and

selecting the first term as a category label for the new image.

2 . The method of claim 1 , wherein processing the new image using the deep neural network to generate the aggregate high-dimensional representation of the new image comprises:

generating the respective scores for each category of the one or more of the plurality of categories by processing the new image using the deep neural network.

3 . The method of claim 2 , wherein processing the new image using the deep neural network comprises:

processing the new image using the deep neural network to obtain a plurality of initial scores;

selecting a pre-determined number of highest initial scores; and

applying a transformation to each of the selected initial scores to generate the respective scores for the plurality of categories.

4 . The method of claim 3 , wherein applying the transformation to each of the selected initial scores comprises normalizing the selected initial scores.

5 . The method of claim 1 , further comprising:

identifying a second term in the vocabulary of terms having a high-dimensional representation that is second closest to the aggregate high-dimensional representation; and

selecting the second term as another category label for the new image.

6 . The method of claim 1 , wherein identifying the first term comprises selecting a term having a high-dimensional representation that has a largest cosine similarity value with the aggregate high-dimensional representation.

7 . The method of claim 1 , wherein obtaining the data comprising a respective high-dimensional representation for each category label of a plurality of category labels that each correspond to a different category of a plurality of categories comprises:

training a machine learning system configured to process each category label of the plurality of category labels to obtain the respective high-dimensional representation of the category label and to associate each category label with the respective high-dimensional representation of the category label.

8 . A system comprising one or more computers and one or more storage devices storing instructions that when executed by the one or more computers cause the one or more computers to perform operations comprising:

obtaining data comprising a respective high-dimensional representation for each category label of a plurality of category labels that each correspond to a different category of a plurality of categories, wherein the high-dimensional representation of the category label is a numeric representation of the category label in a high-dimensional space;

obtaining a new image to be classified;

processing the new image using a deep neural network to generate an aggregate high-dimensional representation of the new image from the high-dimensional representations of the plurality of category labels and respective scores for each category of one or more of the plurality of categories, wherein each of the respective scores represents a respective likelihood that the new image belongs to the corresponding category;

identifying a first term in a vocabulary of terms having a high-dimensional representation that is closest to the aggregate high-dimensional representation; and

selecting the first term as a category label for the new image.

9 . The system of claim 8 , wherein processing the new image using the deep neural network to generate the aggregate high-dimensional representation of the new image comprises:

generating the respective scores for each category of the one or more of the plurality of categories by processing the new image using the deep neural network.

10 . The system of claim 9 , wherein processing the new image using the deep neural network comprises:

processing the new image using the deep neural network to obtain a plurality of initial scores;

selecting a pre-determined number of highest initial scores; and

applying a transformation to each of the selected initial scores to generate the respective scores for the plurality of categories.

11 . The system of claim 10 , wherein applying the transformation to each of the selected initial scores comprises normalizing the selected initial scores.

12 . The system of claim 8 , the operations further comprise:

identifying a second term in the vocabulary of terms having a high-dimensional representation that is second closest to the aggregate high-dimensional representation; and

selecting the second term as another category label for the new image.

13 . The system of claim 8 , wherein identifying the first term comprises selecting a term having a high-dimensional representation that has a largest cosine similarity value with the aggregate high-dimensional representation.

14 . The system of claim 8 , wherein obtaining the data comprising a respective high-dimensional representation for each category label of a plurality of category labels that each correspond to a different category of a plurality of categories comprises:

training a machine learning system configured to process each category label of the plurality of category labels to obtain the respective high-dimensional representation of the category label and to associate each category label with the respective high-dimensional representation of the category label.

15 . A non-transitory computer storage medium encoded with a computer program, the computer program comprising instructions that when executed by one or more computers cause the one or more computers to perform operations comprising:

obtaining data comprising a respective high-dimensional representation for each category label of a plurality of category labels that each correspond to a different category of a plurality of categories, wherein the high-dimensional representation of the category label is a numeric representation of the category label in a high-dimensional space;

obtaining a new image to be classified;

processing the new image using a deep neural network to generate an aggregate high-dimensional representation of the new image from the high-dimensional representations of the plurality of category labels and respective scores for each category of one or more of the plurality of categories, wherein each of the respective scores represents a respective likelihood that the new image belongs to the corresponding category;

identifying a first term in a vocabulary of terms having a high-dimensional representation that is closest to the aggregate high-dimensional representation; and

selecting the first term as a category label for the new image.

16 . The non-transitory computer storage medium of claim 15 , wherein processing the new image using the deep neural network to generate the aggregate high-dimensional representation of the new image comprises:

generating the respective scores for each category of the one or more of the plurality of categories by processing the new image using the deep neural network.

17 . The non-transitory computer storage medium of claim 16 , wherein processing the new image using the deep neural network comprises:

processing the new image using the deep neural network to obtain a plurality of initial scores;

selecting a pre-determined number of highest initial scores; and

applying a transformation to each of the selected initial scores to generate the respective scores for the plurality of categories.

18 . The non-transitory computer storage medium of claim 17 , wherein applying the transformation to each of the selected initial scores comprises normalizing the selected initial scores.

19 . The non-transitory computer storage medium of claim 15 , further comprising:

identifying a second term in the vocabulary of terms having a high-dimensional representation that is second closest to the aggregate high-dimensional representation; and

selecting the second term as another category label for the new image.

20 . The non-transitory computer storage medium of claim 15 , wherein identifying the first term comprises selecting a term having a high-dimensional representation that has a largest cosine similarity value with the aggregate high-dimensional representation.