Assessing response quality to instructions in language models is vital but challenging due to the complexity of human language across different contexts. This complexity often results in ambiguous or inconsistent interpretations, making accurate assessment difficult. To address this issue, we propose a novel Uncertainty-aware Reward Model (URM) that introduces a robust uncertainty estimation for the quality of paired responses based on Bayesian approximation. Trained with preference datasets, our uncertainty-enabled proxy not only scores rewards for responses but also evaluates their inherent uncertainty. Empirical results demonstrate significant benefits of incorporating the proposed proxy into language model training. Our method boosts the instruction following capability of language models by refining data curation for training and improving policy optimization objectives, thereby surpassing existing methods by a large margin on benchmarks such as Vicuna and MT-bench. These findings highlight that our proposed approach substantially advances language model training and paves a new way of harnessing uncertainty within language models.

3D deep learning is a growing field of interest due to the vast amount of information stored in 3D formats. Triangular meshes are an efficient representation for irregular, non-uniform 3D objects. However, meshes are often challenging to annotate due to their high computational complexity. Therefore, it is desirable to train segmentation networks with limited-labeled data. Self-supervised learning (SSL), a form of unsupervised representation learning, is a growing alternative to fully-supervised learning which can decrease the burden of supervision for training. Specifically, contrastive learning (CL), a form of SSL, has recently been explored to solve limited-labeled data tasks. We propose SSL-MeshCNN, a CL method for pre-training CNNs for mesh segmentation. We take inspiration from prior CL frameworks to design a novel CL algorithm specialized for meshes. Our preliminary experiments show promising results in reducing the heavy labeled data requirement needed for mesh segmentation by at least 33%.

Deploying deep visual models can lead to performance drops due to the discrepancies between source and target distributions. Several approaches leverage labeled source data to estimate target domain accuracy, but accessing labeled source data is often prohibitively difficult due to data confidentiality or resource limitations on serving devices. Our work proposes a new framework to estimate model accuracy on unlabeled target data without access to source data. We investigate the feasibility of using pseudo-labels for accuracy estimation and evolve this idea into adopting recent advances in source-free domain adaptation algorithms. Our approach measures the disagreement rate between the source hypothesis and the target pseudo-labeling function, adapted from the source hypothesis. We mitigate the impact of erroneous pseudo-labels that may arise due to a high ideal joint hypothesis risk by employing adaptive adversarial perturbation on the input of the target model. Our proposed source-free framework effectively addresses the challenging distribution shift scenarios and outperforms existing methods requiring source data and labels for training.

When a new problem is given to be solved with high performance, deep learning needs large effort to tune the model hyperparameters including model architectures and training hyperparameters. Many previous works have tried to tune the model hyperparameters automatically, but the algorithms only target on searching either model architectures or training hyperparameters. However, simultaneous optimization of the model architectures and the training hyperparameters works slow and falls into bad local minimums because of the search space enlarged by the high correlation between the two model hyperparameters. In this paper, we propose a novel algorithm to efficiently find the best set of model architectures and training hyperparameters. To efficiently handle the large search space, the proposed algorithm selectively utilizes the given training samples, while limiting the search space by a novel ensemble sampling method. Also, the evaluation time is further reduced by a novel termination mechanism. The accelerated computation of the proposed algorithm is validated by using complex image datasets, which shows the state-of-the-art performance with the 70.9% reduction of computational time.

An integral part of interactive computing environments are systems that have the ability to process information about their users in real-time. In many cases it is desirable to not only recognize a human user but also to extract as much information about the user as possible, such as gender, ethnicity, age, etc. In this paper we present an FPGA implementation of a neural network configured specifically for performing face detection and gender classification in real-time video streams. Our streaming architecture performs the face and gender classification tasks at 30 frames per second on a small sized Virtex-4 FPGA, at accuracy comparable to that of a leading commercial software implementation.

We present an approach to estimate the poses off human heads in natural scenes. The essential features for estimating the head pose are the positions of the prominent facial features relative to the position of the head. We have developed a high-dimensional, randomly sparse representation of a human face using a simplified facial feature model. The representation transforms a raw face image into a vector, representing how well the image matches a large number of randomly-posed and shaped head models. This transformation is designed to collect salient features of the face image that are useful to estimate the pose, while suppressing any irrelevant variations of face appearance. The relation between the sparse representation and the pose is learned using SVR (support vector regression). The sparse representation combined with SVR is shown to estimate the pose more quickly and accurately than SVR applied to raw images.

We propose an approach to accurately detecting two-dimensional (2-D) shapes. The cross section of the shape boundary is modeled as a step function. We first derive a one-dimensional (1-D) optimal step edge operator, which minimizes both the noise power and the mean squared error between the input and the filter output. This operator is found to be the derivative of the double exponential (DODE) function, originally derived by Ben-Arie and Rao. We define an operator for shape detection by extending the DODE filter along the shape's boundary contour. The responses are accumulated at the centroid of the operator to estimate the likelihood of the presence of the given shape. This method of detecting a shape is in fact a natural extension of the task of edge detection at the pixel level to the problem of global contour detection. This simple filtering scheme also provides a tool for a systematic analysis of edge-based shape detection. We investigate how the error is propagated by the shape geometry. We have found that, under general assumptions, the operator is locally linear at the peak of the response. We compute the expected shape of the response and derive some of its statistical properties. This enables us to predict both its localization and detection performance and adjust its parameters according to imaging conditions and given performance specifications. Applications to the problem of vehicle detection in aerial images, human facial feature detection, and contour tracking in video are presented.

We present a comprehensive treatment of 3D object tracking by posing it as a nonlinear state estimation problem. The measurements are derived using the outputs of shape-encoded filters. The nonlinear state estimation is performed by solving the Zakai equation, and we use the branching particle propagation method for computing the solution. The unnormalized conditional density for the solution to the Zakai equation is realized by the weight of the particle. We first sample a set of particles approximating the initial distribution of the state vector conditioned on the observations, where each particle encodes the set of geometric parameters of the object. The weight of the particle represents geometric and temporal fit, which is computed bottom-up from the raw image using a shape-encoded filter. The particles branch so that the mean number of offspring is proportional to the weight. Time update is handled by employing a second-order motion model, combined with local stochastic search to minimize the prediction error. The prediction adjustment suggested by system identification theory is empirically verified to contribute to global stability. The amount of diffusion is effectively adjusted using a Kalman updating of the covariance matrix. WE have successfully applied this method to human head tracking, where we estimate head motion and compute structure using simple head and facial feature models.

We present an approach to tracking human activities in a monocular video. We model the human body by decomposing it into torso and limbs and use simple 3D shapes to approximate them. The limb motions are parametrized by the relative joint angles. The problems of motion tracking and estimation are posed as nonlinear state estimation problems. The measurements are computed using the outputs of 3D shape-encoded filters which extract the boundary gradient information of the body image. The uncertainties of body pose are propagated by a branching particle system. We first sample a set of particles approximating the initial distribution of the state vector conditioned on observations, where each particle encodes the body pose. The posterior density is realized by the weight of the particle, where the weight represents geometric and temporal fit, and computed bottom-up from the raw image using a shape-encoded filter. The particles branch so that the mean number of offspring is proportional to the weight. Applications to both synthetic and real video sequences show the effectiveness of this approach.

We present a new approach for accurate detection of two-dimensional shapes. We first derive an optimal smoothing filter, which minimizes both the noise power and the mean squared error between the input and the filter output. This operator is found to be a derivative of the double exponential (DODE) function. We define an operator for shape detection by extending the DODE filter along the shape's boundary contour. We find that this filtering scheme is equivalent to integrating gradients along the hypothetical shape boundary, but our method turns out to be more robust than conventional edge detection followed by edge grouping. This approach also provides a tool for a systematic analysis of edge-based shape detection. We investigate how the error is propagated by the shape geometry. This enables us to predict both its localization and detection performance. Application to vehicle detection in aerial images and human facial feature detection are provided.