Few-shot training of a neural network

Abstract

A neural network is trained to identify one or more features of an image. The neural network is trained using a small number of original images, from which a plurality of additional images are derived. The additional images generated by rotating and decoding embeddings of the image in a latent space generated by an autoencoder. The images generated by the rotation and decoding exhibit changes to a feature that is in proportion to the amount of rotation.

1 - 27 . (canceled) 
     
     
         28 . A processor, comprising:
 one or more circuits to use one or more neural networks to estimate an orientation of one or more objects within one or more images based, at least in part, on one or more rotated versions of the one or more objects.   
     
     
         29 . The processor of  claim 28 , wherein the one or more rotated versions of the one or more objects contain an embedding that is rotated by an amount. 
     
     
         30 . The processor of  claim 29 , wherein the estimate of the orientation of the one or more objects within the one or more images is based, at least in part, on the rotated embedding exhibiting a change to at least one of the one or more objects in proportion to the amount of rotation of the embedding. 
     
     
         31 . The processor of  claim 28 , wherein the estimate of the orientation of the one or more objects within the one or more images is based, at least in part, on comparing a value predicted in the one or more rotated versions of the one or more objects to a known value associated with one or more objects. 
     
     
         32 . The processor of  claim 28 , wherein the one or more rotated versions of the one or more objects contain an embedding in a latent space that can be rotated while preserving a property of interest. 
     
     
         33 . The processor of  claim 28 , wherein the estimate of the orientation of the one or more objects within the one or more images is based, at least in part, on the one or more rotated versions of the one or more objects containing an encoded orientation that is generated by an encoder. 
     
     
         34 . The processor of  claim 28 , the estimate of the orientation of the one or more objects within the one or more images is based, at least in part, on the one or more rotated versions of the one or more objects containing an encoded orientation that is generated based at least in part on an encoder trained to map input to points in an equivariant latent space. 
     
     
         35 . A system, comprising:
 one or more processors to cause one or more circuits to use one or more neural networks to estimate an orientation of one or more objects within one or more images based, at least in part, on one or more rotated versions of the one or more objects.   
     
     
         36 . The system of  claim 35 , wherein the one or more rotated versions of the one or more objects contain an embedding that is rotated by an amount. 
     
     
         37 . The system of  claim 36 , wherein the estimate of the orientation of the one or more objects within the one or more images is based, at least in part, on the rotated embedding exhibiting a change to at least one of the one or more objects in proportion to the amount of rotation of the embedding. 
     
     
         38 . The system of  claim 35 , wherein the estimate of the orientation of the one or more objects within the one or more images is based, at least in part, on comparing a value predicted in the one or more rotated versions of the one or more objects to a known value associated with one or more objects. 
     
     
         39 . The system of  claim 35 , wherein the one or more rotated versions of the one or more objects contain an embedding in a latent space that can be rotated while preserving a property of interest. 
     
     
         40 . The system of  claim 35 , wherein the estimate of the orientation of the one or more objects within the one or more images is based, at least in part, on the one or more rotated versions of the one or more objects containing an encoded orientation that is generated by an encoder. 
     
     
         41 . The system of  claim 35 , the estimate of the orientation of the one or more objects within the one or more images is based, at least in part, on the one or more rotated versions of the one or more objects containing an encoded orientation that is generated based, at least in part, on an encoder trained to map input to points in an equivariant latent space. 
     
     
         42 . A method, comprising:
 estimating an orientation of one or more objects within one or more images based, at least in part, on one or more rotated versions of the one or more objects.   
     
     
         43 . The method of  claim 42 , wherein the orientation of the one or more objects within the one or more images is based, at least in part, on the one or more rotated versions of the one or more objects contains an embedding that is rotated by an amount and the rotated embedding exhibiting a change to at least one of the one or more objects in proportion to the amount of rotation of the embedding. 
     
     
         44 . The method of  claim 42 , wherein the orientation of the one or more objects within the one or more images is based, at least in part, on comparing a value predicted in the one or more rotated versions of the one or more objects to a known value associated with one or more objects. 
     
     
         45 . The method of  claim 42 , wherein the one or more rotated versions of the one or more objects contain an embedding in a latent space that can be rotated while preserving a property of interest. 
     
     
         46 . The method of  claim 42 , wherein the orientation of the one or more objects within the one or more images is based, at least in part, on the one or more rotated versions of the one or more objects containing an encoded orientation that is generated by an encoder. 
     
     
         47 . The method of  claim 42 , wherein the orientation of the one or more objects within the one or more images is based, at least in part, on the one or more rotated versions of the one or more objects containing an encoded orientation that is generated based, at least in part, on an encoder trained to map input to points in an equivariant latent space.