We provide a probabilistic interpretation of attention and show that the standard dot-product attention in transformers is a special case of Maximum A Posteriori (MAP) inference. The proposed approach suggests the use of Expectation Maximization algorithms for on-line adaptation of key and value model parameters. This approach is useful for cases in which external agents, e.g., annotators, provide inference-time information about the correct values of some tokens, e.g., the semantic category of some pixels, and we need for this new information to propagate to other tokens in a principled manner. We illustrate the approach on an interactive semantic segmentation task in which annotators and models collaborate online to improve annotation efficiency. Using standard benchmarks, we observe that key adaptation boosts model performance ($\sim10\%$ mIoU) in the low feedback regime and value propagation improves model responsiveness in the high feedback regime. A PyTorch layer implementation of our probabilistic attention model is available here: https://github.com/apple/ml-probabilistic-attention.

Mohsen Malmir1 http://mplab.ucsd.edu/~mmalmir/ Karan Sikka1 http://mplab.ucsd.edu/~ksikka/ Deborah Forster1 [email protected] Javier Movellan2 http://www.emotient.com/ Garrison W. Cottrell3 http://cseweb.ucsd.edu/~gary/ 1 Machine Perception Lab. University of California San Diego, San Diego, CA, USA 2 Emotient, Inc. 4435 Eastgate Mall, Suite 320, San Diego, CA, USA 3 Computer Science and Engineering Dept. University of California San Diego, San Diego, CA, USA

Student engagement is a key concept in contemporary education, where it is valued as a goal in its own right. In this paper we explore approaches for automatic recognition of engagement from students' facial expressions. We studied whether human observers can reliably judge engagement from the face; analyzed the signals observers use to make these judgments; and automated the process using machine learning. We found that human observers reliably agree when discriminating low versus high degrees of engagement (Cohen's κ = 0.96). When fine discrimination is required (four distinct levels) the reliability decreases, but is still quite high ( κ = 0.56). Furthermore, we found that engagement labels of 10-second video clips can be reliably predicted from the average labels of their constituent frames (Pearson r=0.85), suggesting that static expressions contain the bulk of the information used by observers. We used machine learning to develop automatic engagement detectors and found that for binary classification (e.g., high engagement versus low engagement), automated engagement detectors perform with comparable accuracy to humans. Finally, we show that both human and automatic engagement judgments correlate with task performance. In our experiment, student post-test performance was predicted with comparable accuracy from engagement labels ( r=0.47) as from pre-test scores ( r=0.44).

In machine learning and computer vision, input signals are often filtered to increase data discriminability. For example, preprocessing face images with Gabor band-pass filters is known to improve performance in expression recognition tasks [1]. Sometimes, however, one may wish to purposely decrease discriminability of one classification task (a “distractor” task), while simultaneously preserving information relevant to another task (the target task): For example, due to privacy concerns, it may be important to mask the identity of persons contained in face images before submitting them to a crowdsourcing site (e.g., Mechanical Turk) when labeling them for certain facial attributes. Suppressing discriminability in distractor tasks may also be needed to improve inter-dataset generalization: training datasets may sometimes contain spurious correlations between a target attribute (e.g., facial expression) and a distractor attribute (e.g., gender). We might improve generalization to new datasets by suppressing the signal related to the distractor task in the training dataset. This can be seen as a special form of supervised regularization. In this paper we present an approach to automatically learning preprocessing filters that suppress discriminability in distractor tasks while preserving it in target tasks. We present promising results in simulated image classification problems and in a realistic expression recognition problem.

As robots find applications in daily life conditions it becomes important to develop controllers that generate energy efficient movements by restricting variability and utilizing high gains only when necessary. Here we present a computationally light and energy efficient approach (AAC) that combines an anticipatory open-loop controller and a variable gain closed loop controller. The approach is grounded in the theory of stochastic optimal control and feedback linearization. As such it links two important approaches to robot control: (1) the family of Computed Torque Controllers (CTC) that are grounded on feedback linearization and classic feedback control, and (2) a more recent family of controllers that aim at finding approximately optimal trade-offs between task performance and energy consumption. Here we show that AAC controllers are highly energy efficient, when compared to CTC, and exhibit some key properties of human motion.

We present a new machine learning algorithm for learning optimal feedback control policies to guide a robot to a goal in the presence of obstacles. Our method works by first reducing the problem of obstacle avoidance to a continuous state, action, and time control problem, and then uses efficient collocation methods to solve for an optimal feedback control policy. This formulation of the obstacle avoidance problem improves over standard approaches, such as potential field methods, by being resistant to local minima, allowing for moving obstacles, handling stochastic systems, and computing feedback control strategies that take into account the robot's (possibly non-linear) dynamics. In addition to contributing a new method for obstacle avoidance, our work contributes to the state-of-the-art in collocation methods for non-linear stochastic optimal control problems in two important ways: (1) we show that taking into account local gradient and second-order derivative information of the optimal value function at the collocation points allows us to exploit knowledge of the derivative information about the system dynamics, and (2) we show that computational savings can be achieved by directly fitting the gradient of the optimal value function rather than the optimal value function itself. We validate our approach on three problems: non-convex obstacle avoidance of a point-mass robot, obstacle avoidance for a 2 degree of freedom robotic manipulator, and optimal control of a non-linear dynamical system.

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 the 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 non-parametric 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 out-performed previous parametric and non-parametric approaches as well as previous methods for combining the two approaches.

In expression recognition and many other computer vision applications, the recognition performance is greatly improved by adding a layer of nonlinear texture filters between the raw input pixels and the classifier. The function of this layer is typically known as feature extraction. Popular filter types for this layer are Gabor energy filters (GEFs) and local binary patterns (LBPs). Recent work [1] suggests that adding a second layer of nonlinear filters on top of the first layer may be beneficial. However, it is unclear what is the best architecture of layers and selection of filters. In this paper, we present a thorough empirical analysis of the performance of single-layer and dual-layer texture-based approaches for action unit recognition. For the single hidden layer case, GEFs perform consistently better than LBPs, which may be due to their robustness to jitter and illumination noise as well as to their ability to encode texture at multiple resolutions. For dual-layer case, we confirm that, while small, the benefit of adding this second layer is reliable and consistent across data sets. Interestingly for this second layer, LBPs appear to perform better than GEFs.

In this paper, we explore the use of computer vision techniques to analyze students' moods during one-on-one teaching interactions. The eventual goal is to create automated tutoring systems that are sensitive to the student's mood and affective state. We find that the problem of accurately determining a child's mood from a single video frame is surprisingly difficult, even for humans. However when the system is allowed to make decisions based on information from 10 to 30 seconds of video, excellent performance may be obtained.

This paper assesses the performance of measures of facial expression dynamics derived from the Computer Expression Recognition Toolbox (CERT) for classifying emotions in the Facial Expression Recognition and Analysis (FERA) Challenge. The CERT system automatically estimates facial action intensity and head position using learned appearance-based models on single frames of video. CERT outputs were used to derive a representation of the intensity and motion in each video, consisting of the extremes of displacement, velocity and acceleration. Using this representation, emotion detectors were trained on the FERA training examples. Experiments on the released portion of the FERA dataset are presented, as well as results on the blind test. No consideration of subject identity was taken into account in the blind test. The F1 scores were well above the baseline criterion for success.

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 prototypical 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 facial 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.

We explore how CERT, 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 are 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.

Face and eye detection algorithms are deployed in a wide variety of applications. Unfortunately, there has been no quantitative comparison of how these detectors perform under difficult circumstances. We created a dataset of low light and long distance images which possess some of the problems encountered by face and eye detectors solving real world problems. The dataset we created is composed of reimaged images (photohead) and semi-synthetic heads imaged under varying conditions of low light, atmospheric blur, and distances of 3m, 50m, 80m, and 200m. This paper analyzes the detection and localization performance of the participating face and eye algorithms compared with the Viola Jones detector and four leading commercial face detectors. Performance is characterized under the different conditions and parameterized by per-image brightness and contrast. In localization accuracy for eyes, the groups/companies focusing on long-range face detection outperform leading commercial applications.

It is important for drowsiness detection systems to identify different levels of drowsiness and respond appropriately at each level. This study explores how to discriminate moderate from acute drowsiness by applying computer vision techniques to the human face. In our previous study, spontaneous facial expressions measured through computer vision techniques were used as an indicator to discriminate alert from acutely drowsy episodes. In this study we are exploring which facial muscle movements are predictive of moderate and acute drowsiness. The effect of temporal dynamics of action units on prediction performances is explored by capturing temporal dynamics using an over complete representation of temporal Gabor Filters. In the final system we perform feature selection to build a classifier that can discriminate moderate drowsy from acute drowsy episodes. The system achieves a classification rate of .96 A' in discriminating moderately drowsy versus acutely drowsy episodes. Moreover the study reveals new information in facial behavior occurring during different stages of drowsiness.

In this paper, we evaluate adaptive sound localization algorithms for robotic heads. To this end we built a 3 degree-of-freedom head with two microphones encased in artificial pinnae (outer ears). The geometry of the head and pinnae induce temporal differences in the sound recorded at each microphone. These differences change with the frequency of the sound, location of the sound, and orientation of the robot in a complex manner. To learn the relationship between these auditory differences and the location of a sound source, we applied machine learning methods to a database of different audio source locations and robot head orientations. Our approach achieves a mean error of 2.5 degrees for azimuth and 11 degrees for elevation for estimating the position of an audio source. The impressive results highlight the benefits of a two-stage regression model to make use of the properties of the artificial pinnae for elevation estimation. In this work, the algorithms were trained using ground truth data provided by a motion capture system. We are currently generalizing the approach so that the training signal is provided online based on a real-time face detection and speech detection system.

Determining whether someone is talking has applications in many areas such as speech recognition, speaker diarization, social robotics, facial expression recognition, and human computer interaction. One popular approach to this problem is audio-visual synchrony detection. A candidate speaker is deemed to be talking if the visual signal around that speaker correlates with the auditory signal. Here we show that with the proper visual features (in this case movements of various facial muscle groups), a very accurate detector of speech can be created that does not use the audio signal at all. Further we show that this person independent visual-only detector can be used to train very accurate audio-based person dependent voice models. The voice model has the advantage of being able to identify when a particular person is speaking even when they are not visible to the camera (e.g. in the case of a mobile robot). Moreover, we show that a simple sensory fusion scheme between the auditory and visual models improves performance on the task of talking detection. The work here provides dramatic evidence about the efficacy of two very different approaches to multimodal speech detection on a challenging database.

We present a generative model and inference algorithm for 3D nonrigid object tracking. The model, which we call G-flow, enables the joint inference of 3D position, orientation, and nonrigid deformations, as well as object texture and background texture. Optimal inference under G-flow reduces to a conditionally Gaussian stochastic filtering problem. The optimal solution to this problem reveals a new space of computer vision algorithms, of which classic approaches such as optic flow and template matching are special cases that are optimal only under special circumstances. We evaluate G-flow on the problem of tracking facial expressions and head motion in 3D from single-camera video. Previously, the lack of realistic video data with ground truth nonrigid position information has hampered the rigorous evaluation of nonrigid tracking. We introduce a practical method of obtaining such ground truth data and present a new face video data set that was created using this technique. Results on this data set show that G-flow is much more robust and accurate than current deterministic optic-flow-based approaches.

Recent years have seen the development of fast and accurate algorithms for detecting objects in images. However, as the size of the scene grows, so do the running-times of these algorithms. If a 128×102 pixel image requires 20 ms to process, searching for objects in a 1280×1024 image will take 2 s. This is unsuitable under real-time operating constraints: by the time a frame has been processed, the object may have moved. An analogous problem occurs when controlling robot camera that need to scan scenes in search of target objects. In this paper, we consider a method for improving the run-time of general-purpose object-detection algorithms. Our method is based on a model of visual search in humans, which schedules eye fixations to maximize the long-term information accrued about the location of the target of interest. The approach can be used to drive robot cameras that physically scan scenes or to improve the scanning speed for very large high resolution images. We consider the latter application in this work by simulating a “digital fovea” and sequentially placing it in various regions of an image in a way that maximizes the expected information gain. We evaluate the approach using the OpenCV version of the Viola-Jones face detector. After accounting for all computational overhead introduced by the fixation controller, the approach doubles the speed of the standard Viola-Jones detector at little cost in accuracy.

We report results of a study in which a low cost sociable robot was immersed at an Early Childhood Education Center for a period of 2 weeks. The study was designed to investigate whether the robot, which operated fully autonomously during the intervention period, could improve target vocabulary skills of 18-24 month of age toddlers. The results showed a 27 % improvement in knowledge of the target words taught by the robot when compared to a matched set of control words. The results suggest that sociable robots may be an effective and low cost technology to enrich Early Childhood Education environments.

Modern machine learning-based approaches to computer vision require very large databases of labeled images. Some contemporary vision systems already require on the order of millions of images for training (e.g., Omron face detector). While the collection of these large databases is becoming a bottleneck, new Internet-based services that allow labelers from around the world to be easily hired and managed provide a promising solution. However, using these services to label large databases brings with it new theoretical and practical challenges: (1) The labelers may have wide ranging levels of expertise which are unknown a priori, and in some cases may be adversarial; (2) images may vary in their level of difficulty; and (3) multiple labels for the same image must be combined to provide an estimate of the actual label of the image. Probabilistic approaches provide a principled way to approach these problems. In this paper we present a probabilistic model and use it to simultaneously infer the label of each image, the expertise of each labeler, and the difficulty of each image. On both simulated and real data, we demonstrate that the model outperforms the commonly used ``Majority Vote heuristic for inferring image labels, and is robust to both adversarial and noisy labelers.

Localizing facial features is a critical component in many computer vision applications such as expression recognition, face recognition, face tracking, animation, and red-eye correction. Practical applications require detectors that operate reliably under a wide range of conditions, including variations in illumination, pose, ethnicity, gender and age. One challenge for the development of such detectors is the inherent trade-off between robustness and precision. Robust detectors tend to provide poor localization and detectors sensitive to small changes in local structure, which are needed for precise localization, generate a large number of false alarms. Here we present an approach to this trade-off based on context dependent inference. First, robust detectors are used to detect contexts in which target features occur, then precise detectors are trained to localize the features given the detected context. This paper describes the approach and presents a thorough empirical examination of the parameters needed to achieve practical levels of performance, including the size of the training database, size of the detector's receptive fields and methods for information integration. The approach operates in real time and achieves, to our knowledge, the most accurate localization performance to date.

Machine learning approaches have produced some of the highest reported performances for facial expression recognition. However, to date, nearly all automatic facial expression recognition research has focused on optimizing performance on a few databases that were collected under controlled lighting conditions on a relatively small number of subjects. This paper explores whether current machine learning methods can be used to develop an expression recognition system that operates reliably in more realistic conditions. We explore the necessary characteristics of the training data set, image registration, feature representation, and machine learning algorithms. A new database, GENKI, is presented which contains pictures, photographed by the subjects themselves, from thousands of different people in many different real-world imaging conditions. Results suggest that human-level expression recognition accuracy in real-life illumination conditions is achievable with machine learning technology. However, the data sets currently used in the automatic expression recognition literature to evaluate progress may be overly constrained and could potentially lead research into locally optimal algorithmic solutions.

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 (Ekman & Friesen, 1978). Head pose (yaw, pitch, and roll) is also detected using an algorithm presented at this conference (Whitehill & Movellan, 2008). A sample output is shown in Figure 1.

Accurately estimating the person's head position and orientation is an important task for a wide range of applications such as driver awareness and human-robot interaction. Over the past two decades, many approaches have been suggested to solve this problem, each with its own advantages and disadvantages. In this paper, we present a probabilistic framework called generalized adaptive viewbased appearance model (GAVAM) which integrates the advantages from three of these approaches: (1) the automatic initialization and stability of static head pose estimation, (2) the relative precision and user-independence of differential registration, and (3) the robustness and bounded drift of keyframe tracking. In our experiments, we show how the GAVAM model can be used to estimate head position and orientation in real-time using a simple monocular camera. Our experiments on two previously published datasets show that the GAVAM framework can accurately track for a long period of time (>2 minutes) with an average accuracy of 3.5deg and 0.75 in with an inertial sensor and a 3D magnetic sensor.

We present a fully automatic approach to learning the personal facial attractiveness preferences of individual users directly from example images. The target application is computer assisted search of partners in online dating services. The proposed approach is based on the use of epsiv-SVMs to learn a regression function that maps low level image features onto attractiveness ratings. We present empirical results based on a dataset of images collected from a large online dating site. Our system achieved correlations of up to 0.45 (Pearson correlation) on the attractiveness predictions for individual users. We show evidence that the approach learned not just a universal sense of attraction shared by multiple users, but capitalized on the preferences of individual subjects. Our results are promising and could already be used to facilitate the personalized search of partners in online dating.

We present a discriminative approach to frame-by-frame head pose tracking that is robust to a wide range of illuminations and facial appearances and that is inherently immune to accuracy drift. Most previous research on head pose tracking has been validated on test datasets spanning only a small (< 20) subjects under controlled illumination conditions on continuous video sequences. In contrast, the system presented in this paper was both trained and tested on a much larger database, GENKI, spanning tens of thousands of different subjects, illuminations, and geographical locations from images on the Web. Our pose estimator achieves accuracy of 5.82deg, 5.65deg, and 2.96deg root-mean-square (RMS) error for yaw, pitch, and roll, respectively. A set of 4000 images from this dataset, labeled for pose, was collected and released for use by the research community.

Recent years have seen an explosion of research on the computational modeling of human visual attention in task free conditions, i.e., given an image predict where humans are likely to look. This area of research could potentially provide general purpose mechanisms for robots to orient their cameras. One difficulty is that most current models of visual saliency are computationally very expensive and not suited to real time implementations needed for robotic applications. Here we propose a fast approximation to a Bayesian model of visual saliency recently proposed in the literature. The approximation can run in real time on current computers at very little computational cost, leaving plenty of CPU cycles for other tasks. We empirically evaluate the saliency model in the domain of controlling saccades of a camera in social robotics situations. The goal was to orient a camera as quickly as possible toward human faces. We found that this simple general purpose saliency model doubled the success rate of the camera: it captured images of people 70% of the time, when compared to a 35% success rate when the camera was controlled using an open-loop scheme. After 3 saccades (camera movements), the robot was 96% likely to capture at least one person. The results suggest that visual saliency models may provide a useful front end for camera control in robotics applications.

Social robots face the fundamental challenge of detecting and adapting their behavior to the current social mood. For example, robots that assist teachers in early education must choose different behaviors depending on whether the children are crying, laughing, sleeping, or singing songs. Interactive robotic applications require perceptual algorithms that both run in real time and are adaptable to the challenging conditions of daily life. This paper explores a novel approach to auditory mood detection which was born out of our experience immersing social robots in classroom environments. We propose a new set of low-level spectral contrast features that extends a class of features which have proven very successful for object recognition in the modern computer vision literature. Features are selected and combined using machine learning approaches so as to make decisions about the ongoing auditory mood. We demonstrate excellent performance on two standard emotional speech databases (the Berlin Emotional Speech [W. Burkhardt et al., 2005], and the ORATOR dataset [H. Quast, 2001]). In addition we establish strong baseline performance for mood detection on a database collected from a social robot immersed in a classroom of 18-24 months old children [J. Movellan er al., 2007]. This approach operates in real time at little computational cost. It has the potential to greatly enhance the effectiveness of social robots in daily life environments.

Many popular optimization algorithms, like the Levenberg-Marquardt algorithm (LMA), use heuristic-based controllers'' that modulate the behavior of the optimizer during the optimization process. For example, in the LMA a damping parameter is dynamically modified based on a set rules that were developed using various heuristic arguments. Reinforcement learning (RL) is a machine learning approach to learn optimal controllers by examples and thus is an obvious candidate to improve the heuristic-based controllers implicit in the most popular and heavily used optimization algorithms. Improving the performance of off-the-shelf optimizers is particularly important for time-constrained optimization problems. For example the LMA algorithm has become popular for many real-time computer vision problems, including object tracking from video, where only a small amount of time can be allocated to the optimizer on each incoming video frame. Here we show that a popular modern reinforcement learning technique using a very simply state space can dramatically improve the performance of general purpose optimizers, like the LMA. Most surprisingly the controllers learned for a particular domain appear to work very well also on very different optimization domains. For example we used RL methods to train a new controller for the damping parameter of the LMA. This controller was trained on a collection of classic, relatively small, non-linear regression problems. The modified LMA performed better than the standard LMA on these problems. Most surprisingly, it also dramatically outperformed the standard LMA on a difficult large scale computer vision problem for which it had not been trained before. Thus the controller appeared to have extracted control rules that were not just domain specific but generalized across a wide range of optimization domains."

In this paper, we present a barebones robot which is capable of interacting with humans based on social contingency. It expands the previous work of a contingency detector into having both human-model updating (developmental capability) and policy improvement (learning capability) based on the framework of Infomax control. The proposed new controller interacts with humans in both active and responsive ways handling the turn-taking between them.

This letter presents an analysis of the contrastive divergence (CD) learning algorithm when applied to continuous-time linear stochastic neural networks. For this case, powerful techniques exist that allow a detailed analysis of the behavior of CD. The analysis shows that CD converges to maximum likelihood solutions only when the network structure is such that it can match the first moments of the desired distribution. Otherwise, CD can converge to solutions arbitrarily different from the log-likelihood solutions, or they can even diverge. This result suggests the need to improve our theoretical understanding of the conditions under which CD is expected to be well behaved and the conditions under which it may fail. In, addition the results point to practical ideas on how to improve the performance of CD.

The goal of the RUBI project is to accelerate progress in the development of social robots by addressing the problem at multiple levels, including the development of a scientific agenda, research methods, formal approaches, software, and hardware. The project is based on the idea that progress will go hand-in-hand with the emergence of a new scientific discipline that focuses on understanding the organization of adaptive behavior in real-time within the environments in which organisms operate. As such, the RUBI project emphasizes the process of design by immersion, i.e., embedding scientists, engineers and robots in everyday life environments so as to have these environments shape the hardware, software, and scientific questions as early as possible in the development process. The focus of the project so far has been on social robots that interact with 18 to 24 month old toddlers as part of their daily activities at the Early Childhood Education Center at the University of California, San Diego. In this document we present an overall assessment of the lessons and progress through year two of the project.

The design and development of social robots that interact and assist people in daily life requires moving into unconstrained daily-life environments. This presents unexplored methodological challenges to robotic researchers. Is it possible, for example, to perform useful experiments in the uncontrolled conditions of everyday life environments? How long do these studies need to be to provide reliable results? What evaluations methods can be used?In this paper we present preliminary results on a study designed to evaluate an algorithm for social robots in relatively uncontrolled, daily life conditions. The study was conducted as part of the RUBI project, whose goal is to design and develop social robots by immersion in the environment in which the robots are supposed to operate. First we found that in spite of the relative chaotic conditions and lack of control existing in the daily activities of a child-care center, it is possible to perform experiments in a relatively short period of time and with reliable results. We found that continuous audience response methods borrowed from marketing research provided good inter-observer reliabilities, in the order of 70%, and temporal resolution (the cut-off frequency is in the order of 1 cycle per minute) at low cost (evaluation is performed continuously in real time). We also experimented with objective behavioral descriptions, like tracking children's movement across a room. These approaches complemented each other and provided a useful picture of the temporal dynamics of the child-robot interaction, allowing us to gather baseline data for evaluating future systems. Finally, we also touch the ongoing study of behavior analysis through 3 months long-term child-robot interaction.

We have been conducting HRI (human-robot interaction) studies with the basic principle of design by immersion, which suggests the importance of researchers moving themselves into unconstrained daily-life environments. This is crucial for design and development of social robots that interact and assist people in the daily real-world. In this paper we report findings on a study where a small humanoid robot kept attending at a nursery school on a daily basis for more than three months. We focus on children's touch behavior on the robot, and conduct video analyses based on six categories related to the touch behavior. Results tell us important conditions for designing every-day robots.

We present a systematic comparison of machine learning methods applied to the problem of fully automatic recognition of facial expressions. We report results on a series of experiments comparing recognition engines, including AdaBoost, support vector machines, linear discriminant analysis. We also explored feature selection techniques, including the use of AdaBoost for feature selection prior to classification by SVM or LDA. Best results were obtained by selecting a subset of Gabor filters using AdaBoost followed by classification with support vector machines. The system operates in real-time, and obtained 93% correct generalization to novel subjects for a 7-way forced choice on the Cohn-Kanade expression dataset. The outputs of the classifiers change smoothly as a function of time and thus can be used to measure facial expression dynamics. We applied the system to to fully automated recognition of facial actions (FACS). The present system classifies 17 action units, whether they occur singly or in combination with other actions, with a mean accuracy of 94.8%. We present preliminary results for applying this system to spontaneous facial expressions.

This paper reports progress in the development of a humanoid robot designed for real-time face-to-face interaction with humans. An essential component for face-to-face interaction with humans is being able to find faces, tracking them, and smoothly moving back and forth between gazing to a face and other objects of interest. In this paper we propose a system that integrates peripheral vision and foveal vision in a principled manner using particle filters. The developed system generates hypotheses about face position by using peripheral vision and verifies them by integrating peripheral vision and foveal vision. Even though a face may not be present in foveal vision, while the robot is gazing at another object, it keeps plausible hypotheses about the location of the human face with peripheral vision and restarts the face-following by verifying the hypothesis later. Moreover, this fundamental function provides more rich sensory information for human-robot interaction, such as a human's facial expression and lip motions during utterances.

We present a generative model and stochastic filtering algorithm for si- multaneous tracking of 3D position and orientation, non-rigid motion, object texture, and background texture using a single camera. We show that the solution to this problem is formally equivalent to stochastic fil- tering of conditionally Gaussian processes, a problem for which well known approaches exist [3, 8]. We propose an approach based on Monte Carlo sampling of the nonlinear component of the process (object mo- tion) and exact filtering of the object and background textures given the sampled motion. The smoothness of image sequences in time and space is exploited by using Laplace's method to generate proposal distributions for importance sampling [7]. The resulting inference algorithm encom- passes both optic flow and template-based tracking as special cases, and elucidates the conditions under which these methods are optimal. We demonstrate an application of the system to 3D non-rigid face tracking.

We present ongoing work on a project for automatic recognition of spon- taneous facial actions. Spontaneous facial expressions differ substan- tially from posed expressions, similar to how continuous, spontaneous speech differs from isolated words produced on command. Previous methods for automatic facial expression recognition assumed images were collected in controlled environments in which the subjects delib- erately faced the camera. Since people often nod or turn their heads, automatic recognition of spontaneous facial behavior requires methods for handling out-of-image-plane head rotations. Here we explore an ap- proach based on 3-D warping of images into canonical views. We eval- uated the performance of the approach as a front-end for a spontaneous expression recognition system using support vector machines and hidden Markov models. This system employed general purpose learning mech- anisms that can be applied to recognition of any facial movement. The system was tested for recognition of a set of facial actions defined by the Facial Action Coding System (FACS). We showed that 3D tracking and warping followed by machine learning techniques directly applied to the warped images, is a viable and promising technology for automatic facial expression recognition. One exciting aspect of the approach pre- sented here is that information about movement dynamics emerged out of filters which were derived from the statistics of images.

We present a Monte Carlo approach for training partially observable diffusion processes. We apply the approach to diffusion networks, a stochastic version of continuous recurrent neural networks. The approach is aimed at learning probability distributions of continuous paths, not just expected values. Interestingly, the relevant activation statistics used by the learning rule presented here are inner products in the Hilbert space of square integrable functions. These inner products can be computed using Hebbian operations and do not require backpropagation of error signals. Moreover, standard kernel methods could potentially be applied to compute such inner products. We propose that the main reason that recurrent neural networks have not worked well in engineering applications (e.g., speech recognition) is that they implicitly rely on a very simplistic likelihood model. The diffusion network approach proposed here is much richer and may open new avenues for applications of recurrent neural networks. We present some analysis and simulations to support this view. Very encouraging results were obtained on a visual speech recognition task in which neural networks outperformed hidden Markov models.

A number of current face recognition algorithms use face representations found by unsupervised statistical methods. Typically these methods find a set of basis images and represent faces as a linear combination of those images. Principal component analysis (PCA) is a popular example of such methods. The basis images found by PCA depend only on pairwise relationships between pixels in the image database. In a task such as face recognition, in which important information may be contained in the high-order relationships among pixels, it seems reasonable to expect that better basis images may be found by methods sensitive to these high-order statistics. Independent component analysis (ICA), a generalization of PCA, is one such method. We used a version of ICA derived from the principle of optimal information transfer through sigmoidal neurons. ICA was performed on face images in the FERET database under two different architectures, one which treated the images as random variables and the pixels as outcomes, and a second which treated the pixels as random variables and the images as outcomes. The first architecture found spatially local basis images for the faces. The second architecture produced a factorial face code. Both ICA representations were superior to representations based on PCA for recognizing faces across days and changes in expression. A classifier that combined the two ICA representations gave the best performance.

We examine eight different techniques for developing visual rep(cid:173) resentations in machine vision tasks. In particular we compare different versions of principal component and independent com(cid:173) ponent analysis in combination with stepwise regression methods for variable selection. We found that local methods, based on the statistics of image patches, consistently outperformed global meth(cid:173) ods based on the statistics of entire images. This result is consistent with previous work on emotion and facial expression recognition. In addition, the use of a stepwise regression technique for selecting variables and regions of interest substantially boosted performance.

This paper explores a framework for recognition of image sequences using partially observable stochastic differential equation (SDE) models. Monte-Carlo importance sampling techniques are used for efficient estimation of sequence likelihoods and sequence likelihood gradients. Once the network dynamics are learned, we apply the SDE models to sequence recognition tasks in a manner similar to the way Hidden Markov models (HMMs) are commonly applied. The potential advantage of SDEs over HMMS is the use of contin(cid:173) uous state dynamics. We present encouraging results for a video sequence recognition task in which SDE models provided excellent performance when compared to hidden Markov models.

People are active experimenters, not just passive observers, constantly seeking new information relevant to their goals. A reasonable approach to active information gathering is to ask questions and conduct experiments that maximize the expected information gain, given current beliefs (Fedorov 1972, MacKay 1992, Oaksford & Chater 1994). In this paper we present results on an exploratory experiment designed to study people's active information gathering behavior on a concept task (Tenenbaum 2000). The results of the experiment are analyzed in terms of the expected information gain of the questions asked by subjects.

This article analyzes learning in continuous stochastic neural networks defined by stochastic differential equations (SDE). In particular, it studies gradient descent learning rules to train the equilibrium solutions of these networks. A theorem is given that specifies sufficient conditions for the gradient descent learning rules to be local covariance statistics between two random variables: (1) an evaluator that is the same for all the network parameters and (2) a system variable that is independent of the learning objective. While this article focuses on continuous stochastic neural networks, the theorem applies to any other system with Boltzmann-like equilibrium distributions. The generality of the theorem suggests that instead of suppressing noise present in physical devices, a natural alternative is to use it to simplify the credit assignment problem. In deterministic networks, credit assignment requires an evaluation signal that is different for each node in the network. Surprisingly, when noise is not suppressed, all that is needed is an evaluator that is the same for the entire network and a local Hebbian signal. This modularization of signals greatly simplifies hardware and software implementations. The article shows how the theorem applies to four different learning objectives that span supervised, reinforcement, and unsupervised problems: (1) regression, (2) density estimation, (3) risk minimization, and (4) information maximization. Simulations, implementation issues, and implications for computational neuroscience are discussed.

This paper presents ongoing work on a speaker independent visual speech recognition system. The work presented here builds on previous research efforts in this area and explores the potential use of simple hidden Markov models for limited vocabulary, speaker independent visual speech recognition. The task at hand is recognition of the first four English digits, a task with possible applications in car-phone images were modeled as mixtures of independent dialing. The Gaussian distributions, and the temporal dependencies were captured with standard left-to-right hidden Markov models. The results indicate that simple hidden Markov models may be used to successfully recognize relatively unprocessed image sequences. The system achieved performance levels equivalent to untrained humans when asked to recognize the fIrst four English digits.

The gain of a mode in a connectionist network is a multiplicative constant that amplifies or attenuates the net input to the node. The benefits of adaptive gains in back-propagation networks are explored. It is shown that gradient descent with respect to gain greatly increases learning speed by amplifying those directions in weight space that are successfully chosen by gradient descent on weights. Adaptive gains also allow normalization of weight vectors without loss of computational capacity, and the authors suggest a simple modification of the learning rule that automatically achieves weight normalization. A method for creating small hidden layers by making hidden node gains compete according to similarities between nodes in an effect to improve generalization performance is described. Simulations show that this competition method is more effective than the special case of gain decay.>