[Paper Review] Deep Domain Confusion: Maximizing for Domain Invariance

Eric Tzeng, Judy Hoffman+3 more|arXiv (Cornell University)|Dec 10, 2014

Domain Adaptation and Few-Shot Learning24 references2,349 citations

TL;DR

The paper presents a CNN with an adaptation (bottleneck) layer and a domain confusion loss (based on MMD) to learn representations that are both discriminative and domain-invariant, enabling improved supervised and unsupervised visual domain adaptation on the Office dataset.

ABSTRACT

Recent reports suggest that a generic supervised deep CNN model trained on a large-scale dataset reduces, but does not remove, dataset bias on a standard benchmark. Fine-tuning deep models in a new domain can require a significant amount of data, which for many applications is simply not available. We propose a new CNN architecture which introduces an adaptation layer and an additional domain confusion loss, to learn a representation that is both semantically meaningful and domain invariant. We additionally show that a domain confusion metric can be used for model selection to determine the dimension of an adaptation layer and the best position for the layer in the CNN architecture. Our proposed adaptation method offers empirical performance which exceeds previously published results on a standard benchmark visual domain adaptation task.

Motivation & Objective

• Address dataset bias in visual recognition when training and test data come from different domains.
• Learn a deep representation that is simultaneously discriminative and domain-invariant.
• Enable both supervised and unsupervised domain adaptation within a single CNN framework.

Proposed method

• Introduce an adaptation bottleneck layer within a pre-trained CNN architecture.
• Incorporate a domain confusion loss based on Maximum Mean Discrepancy (MMD) to minimize source-target distribution distance.
• Jointly optimize classification loss on labeled data and MMD-based domain-invariance loss.
• Use MMD to guide where to place the adaptation layer (depth) and its width (dimensionality).
• Train with a forked network: one branch for supervised learning on labeled data, another branch for computing domain confusion on all data.

Experimental results

Research questions

• RQ1 Can a domain confusion objective improve transfer of deep features across visual domains?
• RQ2 Does MMD-guided placement and sizing of an adaptation layer yield better domain-invariant representations than fixed representations?
• RQ3 How does the proposed approach perform under supervised and unsupervised domain adaptation settings on the Office dataset?

Key findings

• The proposed method significantly outperforms prior domain adaptation methods on the Office benchmark in both supervised and unsupervised settings.
• MMD-based layer selection (depth) and dimension selection (width) correctly identify representations that minimize domain discrepancy and maximize target accuracy.
• Under Amazon to Webcam transfer, the method achieves 84.1% ( supervised ) and 60% ( unsupervised ) baseline improvements over prior work, with the proposed approach reaching 91.9% average in Table 1 and 81.2% average in Table 2.
• Regularization with the domain confusion loss helps prevent overfitting to the source domain during fine-tuning, leading to higher final accuracy.
• t-SNE visualizations show domain-mixed, well-separated class clusters in the learned representation, indicating effective domain invariance.