Methods and systems for cross-domain few-shot classification

A system for cross-domain few-shot classification through training of an autoencoder comprising an encoder and a decoder of a deep neural network, the system comprising:

• a memory storing instructions;
• one or more processors coupled to the memory and configured to execute the instructions to:
• receive training datasets comprising a source dataset and a novel dataset, each dataset comprising a plurality of data samples, each data sample being labelled with a class label;
• train the autoencoder using a first training stage, for each training iteration:
• sample a batch of data samples from the source dataset and feed the batch to the encoder to generate a plurality of source feature maps;
• feed the plurality of source feature maps to a fully connected layer to predict the class label of each source feature map of the plurality of source feature maps, then calculate a source classification loss;
• feed the plurality of source feature maps to the decoder to reconstruct the respective data sample of each source feature map to generate a source reconstructed dataset, then calculate a source reconstruction loss;
• and
• calculate a first training stage loss based on the source classification loss and the source reconstruction loss;
• and
• update parameters of the autoencoder using the first training stage loss;
• train the autoencoder using a second training stage, for each training iteration:
• split the novel dataset into a support set and a query set;
• feed the support set to the encoder to generate a plurality of support set feature maps used to determine a prototype for each class label in the plurality of support set feature maps;
• feed the query set to the encoder to generate a plurality of query set feature maps, then calculate a query set metric classification loss, which is a second training stage loss;
• and
• update the parameters of the autoencoder using the second training stage loss;
• and
• use the trained autoencoder to perform an inference task that comprises a cross-domain few-shot classification task.

A non-transitory computer-readable medium having instructions tangibly stored thereon for cross-domain few-shot classification through training of an autoencoder comprising an encoder and a decoder of a deep neural network, wherein the instructions, when executed by a processing unit, causes the processing unit to:

• receive training datasets comprising a source dataset and a novel dataset, each dataset comprising a plurality of data samples, each data sample being labelled with a class label;
• train the autoencoder using a first training stage, for each training iteration:
• sample a batch of data samples from the source dataset and feed the batch to the encoder to generate a plurality of source feature maps;
• feed the plurality of source feature maps to a fully connected layer to predict the class label of each source feature map of the plurality of source feature maps, then calculate a source classification loss;
• feed the plurality of source feature maps to the decoder to reconstruct the respective data sample of each source feature map to generate a source reconstructed dataset, then calculate a source reconstruction loss;
• calculate a first training stage loss based on the source classification loss and the source reconstruction loss;
• and
• update parameters of the autoencoder using the first training stage loss;
• train the autoencoder using a second training stage, for each training iteration:
• split the novel dataset into a support set and a query set;
• feed the support set to the encoder to generate a plurality of support set feature maps used to determine a prototype for each class label in the plurality of support set feature maps;
• feed the query set to the encoder to generate a plurality of query set feature maps, then calculate a query set metric classification loss, which is a second training stage loss;
• and
• update the parameters of the autoencoder using the second training stage loss;
• and
• use the trained autoencoder to perform an inference task that comprises a cross-domain few-shot classification task.