PaperlayerView exact revision on arXiv

Source-linked AI summary

Deep Transfer Learning for Cross-domain Activity Recognition

Jindong Wang, Vincent W. Zheng, Yiqiang Chen, Meiyu Huang

arXiv:1807.07963v2eess.IVstat.ML

TL;DR

Cross-domain activity recognition must address missing labels on some body parts, while choosing useful sources is difficult and dissimilar sources can risk negative transfer. The paper proposes label-independent source selection and an end-to-end transfer network, reporting effective source selection and high activity-transfer accuracy on public datasets.

  • Problem

    Cross-domain activity recognition must infer activities for a body part with missing labels, despite differing signals and the risk that dissimilar sources cause negative transfer.

    Source: §1 · p. 1, KEYWORDS · p. 1

  • Method

    USSAR selects source domains without target labels using general and activity-specific relationships, followed by TNNAR, an end-to-end network for classification and domain adaptation.

    Source: KEYWORDS · p. 2, §3.2 · p. 3

  • Results

    USSAR effectively selects source domains and TNNAR achieves the best activity-transfer accuracy, with a 3.42% average improvement over the best method STL.

    Source: KEYWORDS · p. 2, §4.3 · p. 6

  • Takeaways & Limitations

    The findings support selecting related sources for activity transfer rather than relying on all body-part domains, which may include dissimilar sources.

    Source: KEYWORDS · pp. 1–2

    Deep-learning-based cross-domain activity recognition still has open problems.

    Source: §2.1 · p. 3

Abstract

from arXiv · show

Human activity recognition plays an important role in people's daily life. However, it is often expensive and time-consuming to acquire sufficient labeled activity data. To solve this problem, transfer learning leverages the labeled samples from the source domain to annotate the target domain which has few or none labels. Unfortunately, when there are several source domains available, it is difficult to select the right source domains for transfer. The right source domain means that it has the most similar properties with the target domain, thus their similarity is higher, which can facilitate transfer learning. Choosing the right source domain helps the algorithm perform well and prevents the negative transfer. In this paper, we propose an effective Unsupervised Source Selection algorithm for Activity Recognition (USSAR). USSAR is able to select the most similar $K$ source domains from a list of available domains. After this, we propose an effective Transfer Neural Network to perform knowledge transfer for Activity Recognition (TNNAR). TNNAR could capture both the time and spatial relationship between activities while transferring knowledge. Experiments on three public activity recognition datasets demonstrate that: 1) The USSAR algorithm is effective in selecting the best source domains. 2) The TNNAR method can reach high accuracy when performing activity knowledge transfer.

1 INTRODUCTION

Cross-domain activity recognition addresses missing activity labels on one body part by using data from other parts. Human activity recognition also extracts high-level information from sensor inputs and has applications including localization and sleep-state detection.

1 INTRODUCTION

  • Human activity recognition infers high-level information from sensor inputs, such as predicting whether someone is walking or running from smartphone or wristband data.The paper identifies indoor localization, sleep-state detection, and smart-home sensing as applications.Source: §1 · p. 1
  • Cross-domain activity recognition uses information from labeled body parts to build a model for a body part whose activity labels are missing.The paper calls the missing-label body part the target domain and the other body parts source domains.Source: §1 · p. 1
  • The paper is classified under ubiquitous computing and transfer learning.Source: §1 · p. 1

KEYWORDS

The paper addresses cross-domain activity recognition when body-part signals differ and labels are missing. It proposes label-independent source selection and an end-to-end transfer network, then reports effective source selection and high transfer accuracy.

KEYWORDS

  • When body-part labels are missing, cross-domain activity recognition is challenging because signals differ, target labels are unavailable, and dissimilar sources may cause negative transfer.Similar body parts can also occur across different people.Source: KEYWORDS · pp. 1–2
  • USSAR selects source domains without target labels by considering general signal distances and activity-specific relationships between body parts.It aims to capture both generality and specificity in activity information.Source: KEYWORDS · p. 2
  • TNNAR performs classification and adaptation end-to-end on original activity data, using convolutional blocks, an LSTM, and jointly optimized adaptation and classification losses.The convolutional layers capture spatial features and the LSTM captures temporal features.Source: KEYWORDS · p. 2
  • The evaluation uses OPPORTUNITY, PAMAP2, and UCI DSADS for cross-domain recognition across different body parts.Source: KEYWORDS · p. 2
  • USSAR is effective in choosing source domains, and TNNAR achieves the best accuracy in activity transfer.Source: KEYWORDS · p. 2
  • The paper presents USSAR as its first unsupervised source-selection algorithm for activity recognition and evaluates both algorithms in cross-position recognition.It states that the algorithms can increase activity models’ generalization ability.Source: KEYWORDS · p. 2

2 RELATED WORK

Related work frames cross-domain activity recognition as a distribution-mismatch problem and contrasts the paper’s domain selection and activity-specific transfer approach with prior methods.

2.1 Activity Recognition

  • Conventional activity-recognition methods typically assume matching training and test distributions, an assumption that can lead to underfitting in cross-domain recognition.Source: §2.1 · p. 2
  • Deep-learning activity recognition can extract high-level features from raw sensor readings, but open problems remain for deep-learning-based cross-domain recognition [39].Source: §2.1 · pp. 2–3

2.2 Transfer Learning

  • Transfer learning is categorized as instance-based, parameter-based, or feature-based, using reweighting, target clustering, or feature transformations, respectively [38].Source: §2.2 · p. 3
  • Prior source-selection work focused on samples or generic multi-source selection, whereas USSAR selects whole domains and targets activity-specific features [10] [35].Source: §2.2 · p. 3
  • Deep transfer methods such as DDC, JAN, and DANN add adaptation layers to reduce domain divergence; TNNAR additionally targets temporal and spatial activity relationships [14] [26] [37].Source: §2.2 · p. 3

3 OUR METHOD

The method selects K labeled source domains for an unlabeled target, then transfers knowledge to predict target activities. It combines a label-free domain distance with a neural network designed to model spatial and temporal activity structure.

3.1 Problem definition

  • The problem is to select K<M labeled source domains with different distributions from the unlabeled target, then transfer learning to obtain its activity labels.Source: §3.1 · p. 3

3.2 Unsupervised Source Selection

  • USSAR combines classifier-based A-distance with specific distance built from semantic spatial weights and cosine-based kinetic similarity.The semantic weights are assigned using human experience.Source: §3.2 · pp. 3–4
  • The label-free Context Activity Distance ranks sources, then a greedy procedure adds a source when its distance to the selected set is smaller than its distance to the target.Source: §3.2 · p. 4

3.3 Transfer Neural Network

  • TNNAR uses convolutional layers for spatial features, an LSTM for temporal relationships, and fully connected layers for classification.Source: §3.3 · p. 4
  • An adaptation layer after the first fully connected layer uses MMD to reduce domain discrepancy, jointly optimizing adaptation and classification losses.The authors place it there because higher-layer features are more specific, while lower layers extract common features.Source: §3.3 · pp. 4–5

4 EXPERIMENTAL EVALUATION

Experiments evaluate source selection and transfer across three public activity-recognition datasets and 22 tasks. USSAR selects suitable sources, while TNNAR improves cross-domain recognition and shows robustness to parameter choices.

4.1 Datasets and Setup

  • Evaluation spans OPPORTUNITY, PAMAP2, and DSADS across 22 tasks covering three similarity scenarios; the different-body-parts, different-person scenario was excluded.Experiments use all three sensor types for each body part.Source: §4.1 · p. 5

4.2 Evaluation of USSAR

  • Without ground-truth source-domain distances, downstream classification accuracy serves as the selection-quality measure; results indicate USSAR selects suitable source domains.The evaluation compares source-selection algorithms, including A-distance, SDM, GR, and MMD.Source: §4.2 · pp. 5–6, Figure 4 · p. 6

4.3 Evaluation of TNNAR

  • TNNAR improves cross-domain activity recognition by an average 3.42% compared with the best method, STL, and outperforms comparison methods at all similarity levels.The comparison covers single-source and multiple-source transfer experiments.Source: §4.3 · p. 6, Table 2 · p. 7, Table 3 · p. 7
  • TNNAR combines convolutional and LSTM components to learn spatial and temporal activity information, automatically extract features, and share hyperparameters across domains.The comparison methods use features extracted according to human knowledge, whereas TNNAR takes the original signal as input.Source: §4.3 · pp. 6–7

4.4 Sensitivity Analysis

  • Across λ and µ values from 0 to 1 in one Torso-target task with K=3, results indicate broad parameter robustness and no need for cherry-picking.Classification accuracy is recorded for each parameter choice.Source: Figure 5 · p. 7, §4.4 · p. 7

5 CONCLUSIONS AND FUTURE WORK

The paper presents USSAR for selecting related source domains and TNNAR for transferring activity representations. Experiments support their effectiveness, and future work will extend both algorithms to heterogeneous and distant activity domains.

5 CONCLUSIONS AND FUTURE WORK

  • USSAR considers semantic and kinetic relationships between body parts to select source domains related to the target.Source: §5 · p. 7
  • TNNAR learns transferable activity representations end to end, and experiments report the best classification accuracy against comparison algorithms.Source: §5 · p. 7
  • Future work will extend USSAR and TNNAR to heterogeneous and distant activity domains.Source: §5 · p. 7
Loading 1807.07963v2…