Filters








51 Hits in 0.85 sec

Bias-Aware Heapified Policy for Active Learning [article]

Wen-Yen Chang and Wen-Huan Chiang and Shao-Hao Lu and Tingfan Wu and Min Sun
2019 arXiv   pre-print
The data efficiency of learning-based algorithms is more and more important since high-quality and clean data is expensive as well as hard to collect. In order to achieve high model performance with the least number of samples, active learning is a technique that queries the most important subset of data from the original dataset. In active learning domain, one of the mainstream research is the heuristic uncertainty-based method which is useful for the learning-based system. Recently, a few
more » ... s propose to apply policy reinforcement learning (PRL) for querying important data. It seems more general than heuristic uncertainty-based method owing that PRL method depends on data feature which is reliable than human prior. However, there have two problems - sample inefficiency of policy learning and overconfidence, when applying PRL on active learning. To be more precise, sample inefficiency of policy learning occurs when sampling within a large action space, in the meanwhile, class imbalance can lead to the overconfidence. In this paper, we propose a bias-aware policy network called Heapified Active Learning (HAL), which prevents overconfidence, and improves sample efficiency of policy learning by heapified structure without ignoring global inforamtion(overview of the whole unlabeled set). In our experiment, HAL outperforms other baseline methods on MNIST dataset and duplicated MNIST. Last but not least, we investigate the generalization of the HAL policy learned on MNIST dataset by directly applying it on MNIST-M. We show that the agent can generalize and outperform directly-learned policy under constrained labeled sets.
arXiv:1911.07574v1 fatcat:ct3luwds6jhtzp7vfrpb6ctphu

Leveraging Motion Priors in Videos for Improving Human Segmentation [article]

Yu-Ting Chen, Wen-Yen Chang, Hai-Lun Lu, Tingfan Wu, Min Sun
2018 arXiv   pre-print
Despite many advances in deep-learning based semantic segmentation, performance drop due to distribution mismatch is often encountered in the real world. Recently, a few domain adaptation and active learning approaches have been proposed to mitigate the performance drop. However, very little attention has been made toward leveraging information in videos which are naturally captured in most camera systems. In this work, we propose to leverage "motion prior" in videos for improving human
more » ... tion in a weakly-supervised active learning setting. By extracting motion information using optical flow in videos, we can extract candidate foreground motion segments (referred to as motion prior) potentially corresponding to human segments. We propose to learn a memory-network-based policy model to select strong candidate segments (referred to as strong motion prior) through reinforcement learning. The selected segments have high precision and are directly used to finetune the model. In a newly collected surveillance camera dataset and a publicly available UrbanStreet dataset, our proposed method improves the performance of human segmentation across multiple scenes and modalities (i.e., RGB to Infrared (IR)). Last but not least, our method is empirically complementary to existing domain adaptation approaches such that additional performance gain is achieved by combining our weakly-supervised active learning approach with domain adaptation approaches.
arXiv:1807.11436v1 fatcat:jdiwvoqas5hkhohburdpw2wegu

Exploring Bag of Words Architectures in the Facial Expression Domain [chapter]

Karan Sikka, Tingfan Wu, Josh Susskind, Marian Bartlett
2012 Lecture Notes in Computer Science  
Automatic facial expression recognition (AFER) has undergone substantial advancement over the past two decades. This work explores the application of bag of words (BoW), a highly matured approach for object and scene recognition to AFER. We proceed by first highlighting the reasons that makes the task for BoW differ for AFER compared to object and scene recognition. We propose suitable extensions to BoW architecture for the AFER's task. These extensions are able to address some of the
more » ... s of current state of the art appearance-based approaches to AFER. Our BoW architecture is based on the spatial pyramid framework, augmented by multiscale dense SIFT features, and a recently proposed approach for object classification: locality-constrained linear coding and max-pooling. Combining these, we are able to achieve a powerful facial representation that works well even with linear classifiers. We show that a well designed BoW architecture can provide a performance benefit for AFER, and elements of the proposed BoW architecture are empirically evaluated. The proposed BoW approach supersedes previous state of the art results by achieving an average recognition rate of 96% on AFER for two public datasets.
doi:10.1007/978-3-642-33868-7_25 fatcat:n6rxxwley5gsjg2knqdsdqmwdi

Modeling and identification of pneumatic actuators

Yuval Tassa, Tingfan Wu, Javier Movellan, Emanuel Todorov
2013 2013 IEEE International Conference on Mechatronics and Automation  
Pneumatic actuators are mechanically simple and robust, have good energetic properties due to air compressibility, and are relatively cheap. Despite these advantages they are difficult to control -pressure dynamics have typical timescales on the order of 100ms, and this delay can severely cripple simplistic control approaches. The solution is to use a modelbased controller with a good model of the pressure dynamics. Here we present a general parametric model of these dynamics based on both a
more » ... oretical analysis and an empirical study with a humanoid robot. I. INTRODUCTION AND RELATED WORK Pneumatic actuators are attractive for several reasons. They are naturally back-drivable, have low friction, tunable compliance and are very robust. They have a high strengthto-weight ratio -for example a typical cylinder of 5 cm diameter weighing ∼100 grams, running at a standard 85 psi (=590 KPa) above room pressure, produces 1160 Newtons or 260 pounds of force. Furthermore, the mechanical simplicity of pneumatics makes them inexpensive. The central disadvantage or complication, is that they are much slower than electric motors or hydraulics, with dynamic timescales on the order of ∼ 100ms. In order to properly control a pneumatic system, a good model of these dynamics is required. Models of such systems can in general be classified as physical or parametric models. Physical models are constructed from first principles and attempt to conform as closely as possible to the underlying physical system. Parametric models are functions with unknown constants which are found using a curve-fitting procedure. While both types of models can have good predictive properties, the design objectives are different. The physical model attempts to accurately capture all the physical properties, regardless of how important they are for prediction. The design of a parametric model, while focusing on predictive power, must also take into account secondary objectives, like ensuring good convergence and eliminating local minima in the parameter space. Previous work has focused either on precise physical models of pneumatic systems [1][2][3], or on linearized parametric models [4] . In this paper, we first develop a physical model from first principles, and then use this model to guide the design of a non-linear parametric model. The work most closely related to ours is [5] , where quadratic polynomials are used as a basis for the non-linear parametrization. Rather than general polynomials, we use specially crafted functions, chosen to conform to the predictions of the initial physical †
doi:10.1109/icma.2013.6617958 fatcat:ihzq4blcerbj7gzoa4w2ogmtre

Afterload-related reference values for myocardial work indices

Qiancheng Li, Hui Wang, Haiyan Feng, Tingfan Wu, Ying Yang, Dongmei Gao, Lina Sun
2021 Cardiovascular Ultrasound  
Background The novel noninvasive pressure-strain loop (PSL) is a reliable tool that reflects myocardial work (MW). Systolic blood pressure (SBP) is the only independent factor for MW indices. However, afterload-related reference values have not been previously reported. The aim of the present study was to establish reference values for MW parameters by wide range SBP grading. Methods We prospectively selected healthy individuals and subjects with SBP ≥ 140 mmHg at the time of study without
more » ... rdial remodeling. MW parameters were collected and the reference values achieved were grouped by SBP in 10-mmHg. Results Significant differences were noted among the SBP-groups for global work index (GWI) and global constructive work (GCW). The majority of statistical comparisons of the differences in GWI and GCW were significant at each SBP-group. With SBP ranging from 90 to 189 mmHg, the parameters GWI and GCW tended to increase linearly with afterload. Overall, the global wasted work (GWW) tended to rise as SBP was increased, but not all of the differences noted in GWW were significant for each SBP-group. Global work efficiency (GWE) remained stable across all SBP-groups, with the exception of a slight drop noted when it exceeded 160 mmHg. Conclusions The amount of MW but not the work efficiency varied greatly according to the different afterload. This finding cannot be ignored during clinical research or diagnosis and afterload-related reference values are required to make a reasonable judgment on the myocardial function.
doi:10.1186/s12947-021-00253-2 pmid:34167526 fatcat:7hl3yld6lfbq7ha4tbgkm6nusm

Computer Expression Recognition Toolbox

Marian Bartlett, Gwen Littlewort, Tingfan Wu, Javier Movellan
2008 2008 8th IEEE International Conference on Automatic Face & Gesture Recognition  
We present a live demo of the Computer Expression Recognition Toolbox (CERT) developed at University of California, San Diego. CERT measures facial expressions in real-time, and codes them with respect to expressions of basic emotion, as well as over 20 facial actions from the Facial Action Coding System
doi:10.1109/afgr.2008.4813406 dblp:conf/fgr/BartlettLWM08 fatcat:mcq27id2rbb2dotf3fublp3fdi

Insights on Spontaneous Facial Expressions from Automatic Expression Measurement [chapter]

Marian Bartlett, Gwen Littlewort, Esra Vural, Jake Whitehill, Tingfan Wu, Kang Lee, Javier Movellan
2010 Dynamic Faces  
doi:10.7551/mitpress/9780262014533.003.0015 fatcat:rv3jd54ukzgflagtesepziasb4

STAC: Simultaneous tracking and calibration

Tingfan Wu, Yuval Tassa, Vikash Kumar, Javier Movellan, Emanuel Todorov
2013 2013 13th IEEE-RAS International Conference on Humanoid Robots (Humanoids)  
System identification is an essential first step in robotic control. Here we focus on the calibration of kinematic sensors, such as joint angle potentiometers, tendon/actuator extension sensors and motion capture markers, on complex humanoid robots. Manual calibration with protractors and rulers does not scale to complex humanoids like the ones studied here. Classic automatic approaches cross-calibrate multiple sensor systems on the same robot by exploiting their redundancy. However, these
more » ... aches make the strong assumption that the observed joint angles are functions of the sensor measurements plus observation noise. This assumption is too restrictive on modern humanoids where linear actuators and tendons span multiple joints. Here we formulate the calibration problem as a Bayesian inference process on a generative model where hidden joint-angles generate sensor observations. A novel alternating optimization approach is developed to simultaneously track space-time joint angles and calibrate parameters (STAC). Explicit estimation of joint angles makes it possible to calibrate sensors that otherwise cannot be handled by classical approaches, such as tendons wrapping on complicated surfaces and spanning multiple joints. We evaluate STAC to calibrate joint potentiometer, tendon length sensor and motion capture marker positions, on a 38-DoF humanoid robot with 24 optical markers, and a 24 DoF tendon driven hand with 12 markers. We show that STAC can be applied to problems that cannot be handled with classical approaches. In addition we show that for simpler problems STAC is more robust than classical approaches and other probabilistic approaches such as the Extended Kalman Filter.
doi:10.1109/humanoids.2013.7030016 dblp:conf/humanoids/WuTKMT13 fatcat:gankbdixq5fsncvcfxkw3un5we

Semi-parametric Gaussian process for robot system identification

Tingfan Wu, Javier Movellan
2012 2012 IEEE/RSJ International Conference on Intelligent Robots and Systems  
One reason why control of biomimetic robots is so difficult is the fact that we do not have sufficiently accurate mathematical models of their system dynamics. Recent nonparametric machine learning approaches to system identification have shown good promise, outperforming parameterized mathematical models when applied to complex robot system identification problems. Unfortunately, non-parametric methods perform poorly when applied to regions of the state space that are not densely covered by
more » ... training dataset. This problem becomes particularly critical as the state space grows. Parametric methods use the available data very efficiently but, on the flip side, they only provide crude approximations to the actual system dynamics. In practice the systematic deviations between the parametric mathematical model and its physical realization results in control laws that do not take advantage of the compliance and complex dynamics of the robot. Here we present an approach to robot system identification, named Semi-Parametric Gaussian Processes (SGP), that elegantly combines the advantages of parametric and nonparametric approaches. Computer simulations and a physical implementation of an underactuated robot system identification problem show very promising results. We also demonstrate the applicability of SGP to articulated tree-structured robots of arbitrary complexity. In all experiments, SGP significantly outperformed previous parametric and non-parametric approaches as well as previous methods for combining the two approaches.
doi:10.1109/iros.2012.6385977 dblp:conf/iros/WuM12 fatcat:2m26mmiyunfgtd72motfifidcm

The computer expression recognition toolbox (CERT)

Gwen Littlewort, Jacob Whitehill, Tingfan Wu, Ian Fasel, Mark Frank, Javier Movellan, Marian Bartlett
2011 Face and Gesture 2011  
We present the Computer Expression Recognition Toolbox (CERT), a software tool for fully automatic real-time facial expression recognition, and officially release it for free academic use. CERT can automatically code the intensity of 19 different facial actions from the Facial Action Unit Coding System (FACS) and 6 different protoypical facial expressions. It also estimates the locations of 10 facial features as well as the 3-D orientation (yaw, pitch, roll) of the head. On a database of posed
more » ... acial expressions, Extended Cohn-Kanade (CK+ [1]), CERT achieves an average recognition performance (probability of correctness on a two-alternative forced choice (2AFC) task between one positive and one negative example) of 90.1% when analyzing facial actions. On a spontaneous facial expression dataset, CERT achieves an accuracy of nearly 80%. In a standard dual core laptop, CERT can process 320 × 240 video images in real time at approximately 10 frames per second.
doi:10.1109/fg.2011.5771414 dblp:conf/fgr/LittlewortWWFFMB11 fatcat:icbjiu3wx5chnkppkwgcpkiqby

Action unit recognition transfer across datasets

Tingfan Wu, Nicholas J. Butko, Paul Ruvolo, Jacob Whitehill, Marian S. Bartlett, Javier R. Movellan
2011 Face and Gesture 2011  
We explore how CERT [15] , a computer expression recognition toolbox trained on a large dataset of spontaneous facial expressions (FFD07), generalizes to a new, previously unseen dataset (FERA). The experiment was unique in that the authors had no access to the test labels, which were guarded as part of the FERA challenge. We show that without any training or special adaptation to the new database, CERT performs better than a baseline method trained exclusively on that database. Best results
more » ... achieved by retraining CERT with a combination of old and new data. We also found that the FERA dataset may be too small and idiosyncratic to generalize to other datasets. Training on FERA alone produced good results on FERA but very poor results on FFD07. We reflect on the importance of challenges like this for the future of the field, and discuss suggestions for standardization of future challenges.
doi:10.1109/fg.2011.5771369 dblp:conf/fgr/WuBRWBM11 fatcat:vhgsalptxnfhddisytk3y5dguu

Facial expression recognition using Gabor motion energy filters

Tingfan Wu, Marian S. Bartlett, Javier R. Movellan
2010 2010 IEEE Computer Society Conference on Computer Vision and Pattern Recognition - Workshops  
Spatial Gabor energy filters (GE) are one of the most successful approaches to represent facial expressions in computer vision applications, including face recognition and expression analysis. It is well known that these filters approximate the response of complex cells in primary visual cortex. However these neurons are modulated by the temporal, not just spatial, properties of the visual signal. This suggests that spatio-temporal Gabor filters may provide useful representations for
more » ... s that involve video sequences. In this paper we explore Gabor motion energy filters (GME) as a biologically inspired representation for dynamic facial expressions. Experiments on the Cohn-Kanade expression dataset show that GME outperforms GE, particularly on difficult low intensity expression discrimination.
doi:10.1109/cvprw.2010.5543267 dblp:conf/cvpr/WuBM10 fatcat:zjjzfznhabg3jnobjmgs6w244u

Learning to Make Facial Expressions

Tingfan Wu, Nicholas J. Butko, Paul Ruvulo, Marian S. Bartlett, Javier R. Movellan
2009 2009 IEEE 8th International Conference on Development and Learning  
doi:10.1109/devlrn.2009.5175536 fatcat:qe73psghn5g3vpxac26w56niwi

A mesh optimization method using machine learning technique and variational mesh adaptation

Tingfan WU, Xuejun LIU, Wei AN, Zenghui HUANG, Hongqiang LYU
2021 Chinese Journal of Aeronautics  
Computational mesh is an important ingredient that affects the accuracy and efficiency of CFD numerical simulation. In light of the introduced large amount of computational costs for many adaptive mesh methods, moving mesh methods keep the number of nodes and topology of a mesh unchanged and do not increase CFD computational expense. As the state-of-the-art moving mesh method, the variational mesh adaptation approach has been introduced to CFD calculation. However, quickly estimating the flow
more » ... eld on the updated meshes during the iterative algorithm is challenging. A mesh optimization method, which embeds a machine learning regression model into the variational mesh adaptation, is proposed. The regression model captures the mapping between the initial mesh nodes and the flow field, so that the variational method could move mesh nodes iteratively by solving the mesh functional which is built from the estimated flow field on the updated mesh via the regression model. After the optimization, the density of the nodes in the high gradient area increases while the density in the low gradient area decreases. Benchmark examples are first used to verify the feasibility and effectiveness of the proposed method. And then we use the steady subsonic and transonic flows over cylinder and NACA0012 airfoil on unstructured triangular meshes to test our method. Results show that the proposed method significantly improves the accuracy of the local flow features on the adaptive meshes. Our work indicates that the proposed mesh optimization approach is promising for improving the accuracy and efficiency of CFD computation.
doi:10.1016/j.cja.2021.05.018 fatcat:7ftpfzkudrdvflh3edpgsz2ceq

Hebbian learning of visually directed reaching by a robot arm

Yiwen Wang, Tingfan Wu, Garrick Orchard, Piotr Dudek, Michele Rucci, Bertram E. Shi
2009 2009 IEEE Biomedical Circuits and Systems Conference  
We describe a robotic system consisting of an arm and an active vision system learns to align its sensory and motor maps so that it can successfully reach the tip of its arm to touch the point where it is looking. This system uses an unsupervised Hebbian learning algorithm, and learns the alignment by watching its arm waving in front of its eyes. After watching for 25 minutes, the maps are sufficiently well aligned that it can execute the desired behavior. I. 978-1-4244-4918-7/09/$25.00 ©2009 IEEE
doi:10.1109/biocas.2009.5372049 fatcat:cwvgky6edfccnoqljypgjqt5xa
« Previous Showing results 1 — 15 out of 51 results