[Paper Review] Gradual Domain Adaptation in the Wild:When Intermediate Distributions are Absent

Samira Abnar, Rianne van den Berg+4 more|arXiv (Cornell University)|Jun 10, 2021

Domain Adaptation and Few-Shot Learning50 references4 citations

TL;DR

This paper proposes GIFT, a method for gradual domain adaptation in the absence of intermediate distributions by generating virtual samples through linear interpolation of source and target representations in feature space. It demonstrates that GIFT outperforms iterative self-training when intermediate distributions are missing, achieving significant accuracy gains on both synthetic and real-world benchmarks with natural distribution shifts.

ABSTRACT

We focus on the problem of domain adaptation when the goal is shifting the model towards the target distribution, rather than learning domain invariant representations. It has been shown that under the following two assumptions: (a) access to samples from intermediate distributions, and (b) samples being annotated with the amount of change from the source distribution, self-training can be successfully applied on gradually shifted samples to adapt the model toward the target distribution. We hypothesize having (a) is enough to enable iterative self-training to slowly adapt the model to the target distribution, by making use of an implicit curriculum. In the case where (a) does not hold, we observe that iterative self-training falls short. We propose GIFT, a method that creates virtual samples from intermediate distributions by interpolating representations of examples from source and target domains. We evaluate an iterative-self-training method on datasets with natural distribution shifts, and show that when applied on top of other domain adaptation methods, it improves the performance of the model on the target dataset. We run an analysis on a synthetic dataset to show that in the presence of (a) iterative-self-training naturally forms a curriculum of samples. Furthermore, we show that when (a) does not hold, GIFT performs better than iterative self-training.

Motivation & Objective

• To address the challenge of domain adaptation when intermediate distributions between source and target are unavailable.
• To investigate whether iterative self-training can implicitly form a curriculum when intermediate distributions exist.
• To propose a method that enables effective gradual adaptation even when intermediate data is absent.
• To evaluate the effectiveness of curriculum learning in unsupervised domain adaptation under varying distribution shift conditions.

Proposed method

• GIFT generates virtual intermediate samples by linearly interpolating representations between source and target data in the neural network's feature space.
• The interpolation coefficient is gradually increased during training, simulating a curriculum that starts at the source and moves toward the target distribution.
• The method uses a fixed interpolation schedule to control the difficulty of virtual examples, ensuring a smooth transition from source to target.
• It combines virtual samples with self-training, where pseudo-labels are generated and used to fine-tune the model iteratively.
• The approach is evaluated on both synthetic benchmarks and real-world datasets with natural distribution shifts.
• Confidence-based filtering is applied to select high-quality pseudo-labeled examples during iterative training.

Experimental results

Research questions

• RQ1 Can iterative self-training implicitly form a curriculum when intermediate distributions are present?
• RQ2 How does the absence of intermediate distributions affect the performance of iterative self-training in domain adaptation?
• RQ3 Can virtual intermediate samples generated via feature-space interpolation effectively simulate a curriculum when real intermediate data is missing?
• RQ4 How does GIFT compare to iterative self-training in terms of performance and robustness under limited training iterations or low target domain diversity?

Key findings

• When intermediate distributions are available, iterative self-training forms an implicit curriculum and achieves a ~40% performance improvement on a two-moon dataset compared to training only on the source.
• In the absence of intermediate distributions, iterative self-training fails to adapt effectively, resulting in minimal performance gain on the same two-moon task.
• GIFT improves performance by ~20% on the two-moon dataset where no intermediate data exists, demonstrating its effectiveness in bridging distribution gaps.
• On synthetic CIFAR-10 benchmarks with controlled shifts, GIFT consistently outperforms iterative self-training, especially when training iterations are limited.
• GIFT is more effective than iterative self-training when the target distribution lacks diversity, as it ensures a balanced progression of difficulty through virtual samples.
• Both GIFT and iterative self-training exhibit curriculum-like behavior, as measured by increasing model confidence and accuracy on progressively harder subsets of the target distribution.