Learning Adaptive Deep Representations For Few To Medium Shot Image Classification

Author : Arman Afrasiyabi
Publisher :
ISBN 13 :
Total Pages : 0 pages
Book Rating : 4.:/5 (133 download)

DOWNLOAD NOW!

Download or read book Representation Learning for Few-shot Image Classification written by Arman Afrasiyabi and published by . This book was released on 2022 with total page 0 pages. Available in PDF, EPUB and Kindle. Book excerpt: As the current state-of-the-art machine learning algorithms, deep neural networks require many examples to perform well on a learning task. Gathering and annotating many samples requires significant human labor, and it is even impossible in most real-world problems such as biomedical data analysis. Under the computer vision context, few-shot image classification aims at grasping the human ability to learn new concepts with little supervision. In this respect, the general idea is to transfer knowledge from base categories with more supervision to novel classes with few examples. In particular, the current few-shot learning approaches pre-train a model on available base classes to generalize to the novel classes, perhaps with fine-tuning. However, the current model's generalization is limited because of some assumptions in the pre-training and restrictions in the fine-tuning stage. This thesis aims to relax three assumptions of the current few-shot learning models, and we propose representation learning for few-shot image classification. First, freezing a pre-trained model looks inevitable in the fine-tuning stage due to the high possibility of overfitting on a few examples. Unfortunately, transfer learning with a frozen model assumption limits the model capacity since the model is not updated with any knowledge of the novel classes. In contrast to freezing a model, we propose associative alignment that enables fine-tuning and updating the network on novel categories. Specifically, we present two strategies that detect and align the novel classes to the highly related base categories. While the first strategy pushes the distribution of the novel classes to the center of their related base categories, the second strategy performs distribution matching using an adversarial training algorithm. Overall, our associative alignment aims to prevent overfitting and increase the model capacity by refining the model using novel examples and related base samples. Second, the current few-shot learning approaches perform transferring knowledge to distinctive novel classes under the uni-modal assumption, where all the examples of a single class are represented with a single cluster. Instead, we propose a mixture-based feature space learning (MixtFSL) approach to infer a multi-modal representation. While a previous mixture-model-based work of Allen et al. [1] is based on a classical clustering method in a non-differentiable manner, our MixtFSL is a new end-to-end multi-modal and fully differentiable model. MixtFSL captures the multi-modality of base classes without any classical clustering algorithm using a two-stage framework. The first phase is called initial training and aims to learn preliminary mixture representation with a pair of loss functions. Then, the progressive following stage, the second stage, stabilizes the training with a teacher-student kind of training framework using a single loss function. Third, unlike the current few-shot techniques of representing each input example with a single feature at the end of the network, we propose a set feature extractor and matching feature sets that relax the typical single feature-based assumption by reasoning on feature sets. Here, we hypothesize that the single feature assumption is problematic in few-shot image classification since the novel classes are different from pre-trained base classes. To this end, we propose a new deep learning set feature extractor based on the hybrid convolution-attention neural networks. Additionally, we offer three non-parametric set-to-set metrics to infer the class of the given input. This thesis employs several standard benchmarks of few-shot learning literature and network backbones to evaluate our proposed methods.