[Paper Review] Zero-shot Domain Adaptation without Domain Semantic Descriptors

Atsutoshi Kumagai, Tomoharu Iwata|arXiv (Cornell University)|Jul 9, 2018

Domain Adaptation and Few-Shot Learning39 references16 citations

TL;DR

This paper proposes a zero-shot domain adaptation method that infers domain-specific models for unseen domains without relying on semantic descriptors. It uses latent domain vectors—learned from sets of unlabeled feature vectors in the test domain—via a deep set architecture and variational inference, achieving state-of-the-art performance across multiple datasets, including MNIST-r, IoT, and School.

ABSTRACT

We propose a method to infer domain-specific models such as classifiers for unseen domains, from which no data are given in the training phase, without domain semantic descriptors. When training and test distributions are different, standard supervised learning methods perform poorly. Zero-shot domain adaptation attempts to alleviate this problem by inferring models that generalize well to unseen domains by using training data in multiple source domains. Existing methods use observed semantic descriptors characterizing domains such as time information to infer the domain-specific models for the unseen domains. However, it cannot always be assumed that such metadata can be used in real-world applications. The proposed method can infer appropriate domain-specific models without any semantic descriptors by introducing the concept of latent domain vectors, which are latent representations for the domains and are used for inferring the models. The latent domain vector for the unseen domain is inferred from the set of the feature vectors in the corresponding domain, which is given in the testing phase. The domain-specific models consist of two components: the first is for extracting a representation of a feature vector to be predicted, and the second is for inferring model parameters given the latent domain vector. The posterior distributions of the latent domain vectors and the domain-specific models are parametrized by neural networks, and are optimized by maximizing the variational lower bound using stochastic gradient descent. The effectiveness of the proposed method was demonstrated through experiments using one regression and two classification tasks.

Motivation & Objective

• To address the limitation of existing zero-shot domain adaptation methods that require domain semantic descriptors such as time, location, or device information.
• To enable effective model adaptation to unseen domains without any labeled or unlabeled data from those domains during training.
• To develop a method that infers domain-specific models by learning latent representations of domains directly from feature distributions in the test phase.
• To overcome the instability of inferring domain characteristics from a single feature vector by leveraging sets of feature vectors for robust domain representation.

Proposed method

• Introduces latent domain vectors as disentangled, continuous representations of domains, learned without semantic descriptors.
• Uses a deep set architecture to process variable-sized sets of unlabeled feature vectors from the test domain, enabling permutation-invariant input handling.
• Employs two neural networks: one for feature extraction from input samples, and another for predicting model parameters conditioned on the latent domain vector.
• Applies variational inference to jointly estimate posterior distributions over latent domain vectors and model parameters, optimized via stochastic gradient descent on a variational lower bound.
• Parametrizes the posterior distributions of latent domain vectors and model parameters using neural networks, enabling end-to-end learning.
• Leverages a Bayesian framework to improve generalization, especially when source domain data are limited.

Experimental results

Research questions

• RQ1 Can domain-specific models be effectively inferred for unseen domains without any domain semantic descriptors?
• RQ2 Can latent domain vectors derived from sets of feature vectors better represent domain characteristics than single-sample inference?
• RQ3 How does the performance of the proposed method vary with the dimensionality of the latent domain vector?
• RQ4 How does the number of source domains affect the zero-shot generalization capability of the method?

Key findings

• The proposed method achieved state-of-the-art performance on all three datasets: MNIST-r, IoT, and School, outperforming existing methods in all test scenarios.
• On MNIST-r, the proposed method achieved significantly higher accuracy than all baselines, with a margin of over 10 percentage points in some cases.
• For the IoT dataset, the method greatly improved performance on previously unseen domains such as D-Doorbell and 1002-sec camera, while maintaining competitive results on others.
• On the School dataset, the proposed method showed statistically significant improvement in RMSE across all domain coverage rates (90%, 85%, 75%, 50%).
• The posterior distributions of latent domain vectors were consistently ordered according to domain characteristics (e.g., rotation angles), confirming that the method captures meaningful domain structure.
• The method remained robust across different latent dimensionality values (K=2 to K=10), with consistent performance gains over baseline, indicating stability in the Bayesian framework.