Classifying input examples using a comparison set

1. One or more non-transitory computer storage medium encoded with instructions that, when executed by one or more computers, cause the one or more computers to perform operations for classifying a new image using a comparison set of comparison images, the operations comprising:

maintaining a first comparison set, the first comparison set comprising a plurality of comparison images and a respective label vector for each of the plurality of comparison images, each label vector including a respective score for each label in a predetermined set of labels;

receiving a new image;

determining, for each comparison image, a respective attention weight characterizing a level of similarity between the comparison image and the new image by applying a neural network attention mechanism to the new image and to the comparison image; and

generating a respective label score for each label in the predetermined set of labels from, for each of the comparison images, the respective attention weight for the comparison image and the respective label vector for the comparison image, wherein the respective label score for each of the labels represents a likelihood that the label is a correct label for the new image,

wherein the neural network attention mechanism has been trained by:

sampling a label set from a distribution over a plurality of possible label sets;

sampling a training comparison set and a batch of input images using the sampled set of labels; and

training the neural network attention mechanism to minimize an error predicting the labels for the input images in the batch conditioned on the training comparison set.

2. A method performed by one or more computers for classifying a new image using a comparison set of comparison images, the method comprising:

maintaining a first comparison set, the first comparison set comprising a plurality of comparison images and a respective label vector for each of the plurality of comparison images, each label vector including a respective score for each label in a predetermined set of labels;

receiving a new image;

determining, for each comparison image, a respective attention weight characterizing a level of similarity between the comparison image and the new image by applying a neural network attention mechanism to the new image and to the comparison image; and

generating a respective label score for each label in the predetermined set of labels from, for each of the comparison images, the respective attention weight for the comparison image and the respective label vector for the comparison image, wherein the respective label score for each of the labels represents a likelihood that the label is a correct label for the new image,

wherein the neural network attention mechanism has been trained by:

sampling a label set from a distribution over a plurality of possible label sets;

sampling a training comparison set and a batch of input images using the sampled set of labels; and

training the neural network attention mechanism to minimize an error predicting the labels for the input images in the batch conditioned on the training comparison set.

3. The method of claim 2 , wherein each label set in the plurality of possible label sets is a different combination of labels from a complete set of possible labels.

4. The method of claim 2 , wherein each label set in the plurality of possible label sets includes a same number of labels.

5. The method of claim 2 , wherein different label sets in the plurality of possible label sets include different numbers of labels.

6. The method of claim 2 , wherein the distribution uniformly weights all of the possible label sets, and wherein sampling the label set includes randomly sampling the label set using the distribution.

7. The method of claim 2 , wherein sampling the training comparison set and the batch of input images using the sampled set of labels comprises:

for each label in the sampled label set, sampling a set of training comparison images having that label, wherein the set of sampled training comparison images and their corresponding labels for all of the labels in the sampled label set form the training comparison set, and

wherein the batch of input images includes a random sequence of comparison images that each have a respective one of the labels in the sampled label set.

8. The method of claim 2 , wherein the training comparison set is different from the first comparison set.

9. The method of claim 2 , wherein training the neural network attention mechanism comprises adjusting current values of parameters of the neural network attention mechanism to minimize the error using a neural network training technique.

10. The method of claim 9 , wherein the neural network training technique is a gradient descent-based training technique.

11. The method of claim 2 , wherein determining, for each comparison image, a respective attention weight characterizing the level of similarity between the comparison image and the new image by applying the neural network attention mechanism to the new image and to the comparison image comprises:

processing the comparison image using a comparison image neural network to determine a numeric embedding of the comparison image;

processing the new image using a new image neural network to determine a numeric embedding of the new image; and

determining the respective attention weight by determining a similarity metric between the numeric embedding of the comparison image and the numeric embedding of the new image.

12. The method of claim 11 , wherein the similarity metric is a cosine distance.

13. The method of claim 2 , wherein generating the respective label score for each label in the predetermined set of labels comprises:

for each comparison image, multiplying the label vector for the comparison image by the attention weight for the comparison image to generate a weighted label vector for the comparison image, and

summing the weighted label vectors to generate a combined label vector that includes a respective label score for each label in the predetermined set of labels.

14. A system comprising one or more computers and one or more non-transitory computer storage medium encoded with instructions that, when executed by the one or more computers, cause the one or more computers to perform operations for classifying a new image using a comparison set of comparison images, the operations comprising:

maintaining a first comparison set, the first comparison set comprising a plurality of comparison images and a respective label vector for each of the plurality of comparison images, each label vector including a respective score for each label in a predetermined set of labels;

receiving a new image;

determining, for each comparison image, a respective attention weight characterizing a level of similarity between the comparison image and the new image by applying a neural network attention mechanism to the new image and to the comparison image; and

generating a respective label score for each label in the predetermined set of labels from, for each of the comparison images, the respective attention weight for the comparison image and the respective label vector for the comparison image, wherein the respective label score for each of the labels represents a likelihood that the label is a correct label for the new image,

wherein the neural network attention mechanism has been trained by:

sampling a label set from a distribution over a plurality of possible label sets;

sampling a training comparison set and a batch of input images using the sampled set of labels; and

training the neural network attention mechanism to minimize an error predicting the labels for the input images in the batch conditioned on the training comparison set.

15. The system of claim 14 , wherein each label set in the plurality of possible label sets is a different combination of labels from a complete set of possible labels.

16. The system of claim 14 , wherein the distribution uniformly weights all of the possible label sets, and wherein sampling the label set includes randomly sampling the label set using the distribution.

17. The system of claim 14 , wherein sampling the training comparison set and the batch of input images using the sampled set of labels comprises:

for each label in the sampled label set, sampling a set of training comparison images having that label, wherein the set of sampled training comparison images and their corresponding labels for all of the labels in the sampled label set form the training comparison set, and

wherein the batch of input images includes a random sequence of comparison images that have one of the labels in the sampled label set.

18. The system of claim 14 , wherein the training comparison set is different from the first comparison set.

19. The system of claim 14 , wherein training the neural network attention mechanism comprises adjusting current values of parameters of the neural network attention mechanism to minimize the error using a neural network training technique.

20. The system of claim 19 , wherein the neural network training technique is a gradient descent-based training technique.