Zero and Few Shot Learning With Semantic Feature Synthesis and Competitive Learning

Guan, Jiechao; Lu, Zhiwu; Xiang, Tao; Li, Aoxue; Zhao, An; Wen, Ji-Rong

doi:10.1109/tpami.2020.2965534

Cited by 51 publications

(20 citation statements)

References 45 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…The ZSL approaches can be categorized into classifier-based and instance-based methods. Based on the categorization, the CRL model is defined as a zero-shot learning model [4,43], ConSE [6], CMT [44], SAE [45], Embed [46] Synthesizing method GAN+ALE [47], GAN+Softmax [47], CADA-VAE [48], cycle-GAN [49], BPL+LR [50] Classifier-based ZSL method Relationship method SSE [51], AMP [35], SynC [52] Correspondence method SJE [33], LATEM [37], ALE [5], DeViSE [34], ESZSL [36], SP-AEN [53] that uses image representation semantic space and employs an instance-based projection method for model building and inference. The projection method provides insights on the labeled instances of an unseen class by projecting them onto the common feature space or the semantic space where instances and prototypes are compared [1].…”

Section: Zsl Projection Methodsmentioning

confidence: 99%

“…GAN+ALE [47], GAN+Softmax [47] and cycle-GAN [49] rely on various types of generative adversarial network (GAN) model, which samples a random vector, combines that with the unseen class prototype to form the input to the generator, whereas CADA-VAE [48] uses variational auto-encoder for data synthesizing. BPL [50] uses semantic feature synthesis by perturbation approach that directly perturbs the seen class samples onto unseen class prototypes. In the context of the projection method, BPL [50] uses bidirectional projection learning as the part of a competitive learning strategy between seen samples and unseen prototypes.…”

Section: Zsl Projection Methodsmentioning

confidence: 99%

“…BPL [50] uses semantic feature synthesis by perturbation approach that directly perturbs the seen class samples onto unseen class prototypes. In the context of the projection method, BPL [50] uses bidirectional projection learning as the part of a competitive learning strategy between seen samples and unseen prototypes. Synthesizing methods differ from the CRL model in architectural sense, but we compare it with CRL since it is known as one of the ZSL's top-performing models.…”

Section: Zsl Projection Methodsmentioning

confidence: 99%

“…The four ZSL datasets include Animals with Attributes-2 (AWA2), Caltech-UCSD Birds-200 (CUB200), Scene Attribute dataset (SUN_A), and ImageNet-360 (IN360). IN360 includes 360 unique classes present in ImageNet-1K (2010) version, which is different from the source domain dataset, ImageNet-1K (2015) [42,50,63]. In addition, there (6…”

Section: Datasetsmentioning

confidence: 99%

See 3 more Smart Citations

Class Representative Learning for Zero-shot Learning Using Purely Visual Data

Chandrashekar

Lee

2021

SN COMPUT. SCI.

View full text Add to dashboard Cite

The building of robust classifiers with high precision is an important goal. In reality, it is quite challenging to achieve such a goal with the data that are typically noise, sparse, or derived from heterogeneous sources. Thus, a considerable gap exists between a model built with training (seen) data and testing (unseen) data in applications. Recent works, including zero-shot learning (ZSL) and generalized zero-short learning (G-ZSL), have attempted to overcome the apparent gap through transfer learning. However, most of these works are required to build a model using visual input with associated data like semantics, attributes, and textual information. Furthermore, models are made with all of the training data. Thus, these models apply to more generic contexts but do not apply to the specific settings that will eventually be required for real-world applications. In this paper, we propose a novel model named class representative learning (CRL), a class-based classifier designed with the following unique contributions in machine learning: (1) the unique design of a latent feature vector, i.e., class representative, represents the abstract embedding space projects with the features extracted from a deep neural network learned only from input images. (2) Parallel ZSL algorithms with class representative learning; (3) a novel projection-based inferencing method uses the vector space model to reconcile the dominant difference between the seen classes and unseen classes. This study demonstrates the benefit of using the class-based approach with CRs for ZSL and G-ZSL on eight benchmark datasets. Extensive experimental results suggest that our proposed CRL model significantly outperforms the state-of-the-art methods in ZSL/G-ZSL based image classification.

show abstract

Section: Zsl Projection Methodsmentioning

confidence: 99%

Section: Zsl Projection Methodsmentioning

confidence: 99%

Section: Zsl Projection Methodsmentioning

confidence: 99%

Section: Datasetsmentioning

confidence: 99%

See 2 more Smart Citations

Class Representative Learning for Zero-shot Learning Using Purely Visual Data

Chandrashekar

Lee

2021

SN COMPUT. SCI.

View full text Add to dashboard Cite

show abstract

“…Notwithstanding their great success in image segmentation, F-CNNs require thousands of labeled images for training and their performance degrades when only a small number of annotated images are available [14] . Consequently, an improved mechanism is required for F-CNN training that enables the segmentation of a new semantic class based on a limited number of labeled images [15] . Such approaches frequently use transfer learning (TL) to transfer the knowledge from pre-trained models to offer an initialization that is later enhanced with the new data to adapt to the underlying problem.…”

Section: Introductionmentioning

confidence: 99%