One-Shot Adaptation of Supervised Deep Convolutional Models

By Judy Hoffman, Eric Tzeng, Jeff Donahue, Yangqing Jia, Kate Saenko, Trevor Darrell. 2013

82 citations recorded by the source corpus.

Dataset bias remains a significant barrier towards solving real world computer vision tasks. Though deep convolutional networks have proven to be a competitive approach for image classification, a question remains: have these models have solved the dataset bias problem? In general, training or fine-tuning a state-of-the-art deep model on a new domain requires a significant amount of data, which for many applications is simply not available. Transfer of models directly to new domains without adaptation has historically led to poor recognition performance. In this paper, we pose the following question: is a single image dataset, much larger than previously explored for adaptation, comprehensive enough to learn general deep models that may be effectively applied to new image domains? In other words, are deep CNNs trained on large amounts of labeled data as susceptible to dataset bias as previous methods have been shown to be? We show that a generic supervised deep CNN model trained on a large dataset reduces, but does not remove, dataset bias. Furthermore, we propose several methods for adaptation with deep models that are able to operate with little (one example per category) or no labeled domain specific data. Our experiments show that adaptation of deep models on benchmark visual domain adaptation datasets can provide a significant performance boost.

Subfields: Computer Vision and Pattern Recognition, Representation & Self-Supervised Learning

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Computer Vision and Pattern Recognition

In Representation & Self-Supervised Learning

Citations in the corpus

Domain-Adversarial Training of Neural Networks (2015, background): “, 2014) as it is done in some recent DA works (Donahue et al., 2014; Tzeng et al., 2014; Hoffman et al., 2013; Long and Wang, 2015).”
Hierarchical Multiscale Recurrent Neural Networks (2016, background): “A stack of representation layers, learned from the data in a way to optimize the target task, make deep neural networks entertain advantages such as generalization to unseen examples (Hoffman et al., 2013), sharing learned knowledge among multiple tasks, and discovering disentangling factors of vari”
Continual Few-shot Relation Learning via Embedding Space Regularization and Data Augmentation (2022, methodology): “…(Rezende et al., 2016; 162 Triantafillou et al., 2017; Hu et al., 2018); and 163 (iii) algorithm-based methods try to find a more 164 suitable strategy to search for the best hypothesis in 165 the whole hypothesis space (Hoffman et al., 2013; 166 Ravi and Larochelle, 2017; Finn et al., 2017).”
Generalizing from a Few Examples (2019, methodology): “In [60], it uses the lower layers of a pretrained CNN for feature embedding, and learns a linear classifier on the embedded features using Dtrain.”
A Deeper Look at Dataset Bias (2015, methodology): “However, the analysis is generally restricted to controlled cases where the data difference is limited to speci ﬁ c visual domain shift [7,23].”
Deep Reconstruction-Classification Networks for Unsupervised Domain Adaptation (2016, methodology): “They provide signiﬁcantly better performances than the SURF-based features on the Oﬃce dataset [36,37].”
Cross-Domain Image Retrieval with a Dual Attribute-Aware Ranking Network (2015, unknown): “Several domain adaptation methods based on deep learning have been recently proposed [16, 6].”
DASH-N: Joint Hierarchical Domain Adaptation and Feature Learning (2015, background): “In addition, the learned features have been shown to generalize well across different domains [40], [41], making them suitable for domain adaptation.”
Deep Visual Domain Adaptation: A Survey (2018, background): “DA is adopted by shallow methods, whereas deep networks only extract vectorial features and are not helpful for transferring knowledge directly [16, 41, 72, 64, 110].”
Deep domain adaptation for describing people based on fine-grained clothing attributes (2015, background): “Enhanced R-CNN detection pipeline. activation of the NIN fully-connected layer (FC2) as fea- tures.”
DeepCham: Collaborative Edge-Mediated Adaptive Deep Learning for Mobile Object Recognition (2016, methodology): “This entails coordination challenges of users using a different label for an object class and addressing the image redundancy problem of users taking similar photos of the target domain.”
Where to Buy It: Matching Street Clothing Photos in Online Shops (2015, background): “Recently, supervised deep CNNs have proved to be extremely successful for the domain adaptation task [6, 14, 39].”
Learning visual biases from human imagination (2014, methodology): “n datasets suffer from dataset biases that harm cross dataset generalization performance [35,30]. Unfortunately, there is no known method to ﬁx it (although there have been several good ﬁrst attempts [22,33,19]). Since the imaginary classiﬁers are immune to dataset bias (there is no dataset) and ins”
Joint cross-domain classification and subspace learning for unsupervised adaptation (2014, methodology): “We present an algorithm that learns jointly both a low dimensional representation and a reliable classifier by optimizing a trade-off between the sourcetarget similarity and the source training error.”
Learning to Learn: Model Regression Networks for Easy Small Sample Learning (2016, methodology): “, 16-way classification), as is customary in [43].”