papers related few

∆-encoder: an effective sample synthesis method for few-shot object recognition

• datasets: miniImageNet, CIFAR100, CUB, Caltech-256, APY, SUN and AWA2

• how to generate images: learning to extract transferable intra-class deformations between same-class pairs of training examples and using this deformations to generate samples conditioned on few provided examples.

• how to aid classification task: constructing feeding augmented images into a simple linear N-class classifier (one dense layer followed by softmax) over about 1024 samples of each category

• how to evaluate the effectiveness of generated images: using trained generative model to generate thousands of images conditioned on provided few shots and using those generated images to train a simple classifier (training fc layer with convolutional layers of vgg16 or resnet18 pre-trained on Imagenet, another way is to pre-train those convulutional layers by samples from all training categories). Finally, comparing the results with the accuracy of other few-shot classifiers.

DADA: Deep Adversarial Data Augmentation for Extremely Low Data Regime Classification

• how to generate images: using a class-conditional GAN with a novel loss to generate diverse and category-specific images.

• how to aid classification task: coupling the generation process and classification process. Using 2k loss to distinguish fake data from real data, also to predict the category of fake data and real data.

• how to evaluate the effectiveness of generated images: experiments are not conducted in few-shot setting, instead sampling 50-1000 samples of each category to simulate extremely low data regime. comparing the results with other GAN methods, such as Improved-GAN.

FEW-SHOT AUTOREGRESSIVE DENSITY ESTIMATION: TOWARDS LEARNING TO LEARN DISTRIBUTIONS

• how to generate images: modified pixelCNN to achieve few-shot density estimation and extend the model to generate natural images.

• how to aid classification task: None, only density extimation NLL

• how to evaluate the effectiveness of generated images: visualization of the generated images, comparing the results with the NLL of other pixelCNN methods.

The Variational Homoencoder:Learning to learn high capacity generative models from few examples

• how to generate images: modified VAE to produces a hierarchical latent variable model which better utilises latent variables.

• how to aid classification task: classifying an example x the estimation of the expected conditional likelihood under the variational posterior.

• how to evaluate the effectiveness of generated images: one-shot generation, one-shot classification and the value of likelihood.

Distribution Matching in Variational Inference