Distributionally Robust Semi-Supervised Learning for People-Centric Sensing
Authors: Kaixuan Chen, Lina Yao, Dalin Zhang, Xiaojun Chang, Guodong Long, Sen Wang3321-3328
AAAI 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate our model in a variety of people-centric recognition tasks on real-world datasets, including intention recognition, activity recognition, muscular movement recognition and gesture recognition. The experiment results demonstrate that the proposed model outperforms the state-of-the-art methods. Experiments In this section, we evaluate the performance of our proposed method in four challenging people-centric sensing tasks: intention recognition, activity recognition, muscular movement recognition and gesture recognition. |
| Researcher Affiliation | Academia | Kaixuan Chen,1 Lina Yao,1 Dalin Zhang,1 Xiaojun Chang,2 Guodong Long,3 Sen Wang4 1School of Computer Science and Engineering, University of New South Wales, Australia 2Faculty of Information Technology, Monash University, Australia 3Centre for Artificial Intelligence, FEIT, University of Technology Sydney, Australia 4School of Information and Communication Technology, Griffith University, Australia |
| Pseudocode | Yes | Algorithm 1 Training and Optimization |
| Open Source Code | No | No explicit statement providing access to the source code (e.g., a repository link, or mention of supplementary materials containing code) for the described methodology was found. |
| Open Datasets | Yes | Intention Recognition EEG Dataset (Goldberger et al. 2000): The EEG dataset contains 108 subjects executing left/right fist open and close intention tasks... Muscular Movement Recognition EMG Dataset 1: The UCI EMG Dataset in Lower Limb contains 11 subjects with no abnormalities in the knee executing three different exercises for analysis in the behavior associated with the knee muscle... MHEALTH (Banos et al. 2014): This dataset is devised to benchmark human activity recognition methods based on multimodal wearable sensor data... Gesture Recognition Opportunity Gesture (Roggen et al. 2010): This dataset consists of data collected from four subjects by a wide variety of body-worn, object-based and ambient sensors in a realistic manner. |
| Dataset Splits | Yes | Cross-validation is conducted on all the participant subjects to ensure rigorousness. The data of s U is evenly separated into two, one is the unlabeled training set U and the other is used as the test set T. |
| Hardware Specification | Yes | All the experiments are conducted on a Nvidia Titan X Pascal GPU. |
| Software Dependencies | No | No specific software dependencies with version numbers (e.g., Python 3.x, PyTorch 1.x) were mentioned. The paper refers to optimization rules and activation functions, but not software versions. |
| Experiment Setup | Yes | In this work, we use a convolutional autoencoder as the main architecture. The encoder has one convolutional layer, one max-pooling layer and one fully-connected layer. Two decoders use a mirrored architecture with the encoder, including one fully-connected layer, one un-pooling layer and one deconvolutional layer. Each convolutional layer is followed by a rectified linear unit (Re LU) activation and the classification outputs are calculated by the softmax functions. The kernel size of the convolutional layer and the deconvolutional layers is M × 45 and the number of feature maps is 40, where M denotes the number of features of the datasets and the pooling size is 1 × 75. We use stochastic gradient descent with Adam update rule to minimize the loss functions at a learning rate of 1e-4. Dropout regularization with a keep probability of 0.5 is applied before the fully-connected layers. Batch normalization during training is also used to get better performance. Therefore, we set a threshold threa to seek a balance for the min-max game between person-specific discrepancy and discriminativeness. On the other hand, we require rather strong decoders for reconstruction, a threshold threrec is thus set to guarantee the reconstruction performance. |