[Paper Review] One-Shot Adaptation of Supervised Deep Convolutional Models

Judy Hoffman, Eric Tzeng+4 more|arXiv (Cornell University)|Dec 24, 2013

Domain Adaptation and Few-Shot Learning14 citations

TL;DR

This paper investigates whether large-scale supervised deep CNNs trained on diverse datasets can generalize to new domains with minimal adaptation. It proposes one-shot adaptation methods using just one labeled example per class, demonstrating significant performance gains on visual domain adaptation benchmarks, showing that while dataset bias persists, effective adaptation is achievable with minimal data.

ABSTRACT

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.

Motivation & Objective

• Address the challenge of dataset bias in deep learning for real-world computer vision applications.
• Investigate whether large-scale supervised CNNs trained on diverse data still suffer from dataset bias.
• Develop adaptation techniques that require little or no labeled data from the target domain.
• Enable effective transfer of deep models to new image domains using only one labeled example per category.

Proposed method

• Train a generic deep CNN on a large-scale labeled dataset to learn generalizable features.
• Apply domain adaptation techniques that fine-tune the model using only one labeled example per class in the target domain.
• Utilize feature-space adaptation strategies to align source and target domain distributions with minimal supervision.
• Leverage the pre-trained model's hierarchical feature representations to improve zero-shot or one-shot generalization.
• Apply adaptation via metric learning or feature normalization to reduce domain shift.
• Evaluate adaptation performance on standard visual domain adaptation benchmarks with limited target domain annotations.

Experimental results

Research questions

• RQ1 Can a deep CNN trained on a large dataset generalize effectively to a new domain with minimal adaptation?
• RQ2 To what extent does dataset bias persist in large-scale supervised deep models despite extensive pre-training?
• RQ3 Can effective adaptation be achieved with only one labeled example per class in the target domain?
• RQ4 How do proposed adaptation methods compare to standard fine-tuning or transfer learning in low-data regimes?

Key findings

• Large-scale supervised deep CNNs still exhibit dataset bias, indicating that pre-training alone does not fully resolve domain shift.
• One-shot adaptation methods significantly improve recognition accuracy on benchmark visual domain adaptation datasets.
• The proposed adaptation techniques achieve notable performance boosts even with only one labeled example per class in the target domain.
• Adaptation with minimal data outperforms direct transfer or standard fine-tuning in low-resource domain adaptation scenarios.