Data Augmented Meta-Learning

As a new approach in machine learning,meta learning aims to address the challenge that traditional deep learning is difficult to adapt to new tasks quickly.In recent years,meta-learning has made remarkable achievements in several fields.However,current metalearning models are limited by the lack of data and still have room for further improvement.Data augmentation,a technique commonly used in deep learning to improve model performance,can also be applied to meta learning.In this thesis,we will explore more efficient data augmented meta-learning models from the perspective of data augmented meta-learning.The main work of this thesis includes:(1)Prototype Augmentation with Dummy Samples(PADS)is proposed,which combines generative model and Prototypcial Network model and generates dummy samples in the feature space using a modified Variational Auto Encoder.Two samples are combined as the input to the encoder,and then we maximizing the mutual information between the dummy sample generated by the decoder and the original input samples.This allows the dummy sample to contain a mixture of information from both two input samples.PADS uses only the information from support sets to expands the number of samples in support sets,thereby the prototype of each class is augmented.The augmented prototypes are able to better classify samples from different classes in the feature space.(2)This paper proposed Patch Mix Augmentation with Dual Encoders for Meta-Learning(PMADE)using wavelet transform and the idea of AdaIN to mix and augment image samples at patch level.PMADE utilizes Vision Transformer(ViT)as an auxiliary feature encoder to better perform feature extraction on the patch mixed samples.In order to train our ViT model in the case of few-shot classification,PMADE incorporates the idea of contrastive learning and adopts InfoNCE as the loss function for ViT training.With the generated mixed samples,PMADE helps the meta-learning model to perform better sample classification.(3)This paper proposed Salicency Mix Hallucinator for Meta-Learning which can improve the usability of the samples generated by the hallucinator network by using the saliency map to guide the patch mixing processing between samples.The model first obtains the corresponding saliency map for each sample.Then,the mixed samples are obtained from replacing patches in the original images based on the significance maps.The mixed samples are used as noise inputs to the hallucination network.Finally,the model uses these generated samples to add extra data to the support set for each category,to enhance the classification performance of the meta-learning model.