Bayesian Inference offers principled tools to tackle many critical problems with modern neural networks such as poor calibration and generalization, and data inefficiency. However, scaling Bayesian inference to large architectures is challenging and requires restrictive approximations. Monte Carlo Dropout has been widely used as a relatively cheap way for approximate Inference and to estimate uncertainty with deep neural networks. Traditionally, the dropout mask is sampled independently from a fixed distribution. Recent works show that the dropout mask can be viewed as a latent variable, which can be inferred with variational inference. These methods face two important challenges: (a) the posterior distribution over masks can be highly multi-modal which can be difficult to approximate with standard variational inference and (b) it is not trivial to fully utilize sample-dependent information and correlation among dropout masks to improve posterior estimation. In this work, we propose GFlowOut to address these issues. GFlowOut leverages the recently proposed probabilistic framework of Generative Flow Networks (GFlowNets) to learn the posterior distribution over dropout masks. We empirically demonstrate that GFlowOut results in predictive distributions that generalize better to out-of-distribution data, and provide uncertainty estimates which lead to better performance in downstream tasks.
Humans can classify an unseen category by reasoning on its language explanations. This ability is owing to the compositional nature of language: we can combine previously seen concepts to describe the new category. For example, we might describe mavens as "a kind of large birds with black feathers", so that others can use their knowledge of concepts "large birds" and "black feathers" to recognize a maven. Inspired by this observation, in this work we tackle zero-shot classification task by logically parsing and reasoning on natural language explanations. To this end, we propose the framework CLORE (Classification by LOgical Reasoning on Explanations). While previous methods usually regard textual information as implicit features, CLORE parses the explanations into logical structure the and then reasons along this structure on the input to produce a classification score. Experimental results on explanation-based zero-shot classification benchmarks demonstrate that CLORE is superior to baselines, mainly because it performs better on tasks requiring more logical reasoning. Alongside classification decisions, CLORE can provide the logical parsing and reasoning process as a form of rationale. Through empirical analysis we demonstrate that CLORE is also less affected by linguistic biases than baselines.
Hybrid tabular-textual question answering (QA) requires reasoning from heterogeneous information, and the types of reasoning are mainly divided into numerical reasoning and span extraction. Despite being the main challenge of the task compared to extractive QA, current numerical reasoning method simply uses LSTM to autoregressively decode program sequences, and each decoding step produces either an operator or an operand. However, the step-by-step decoding suffers from exposure bias, and the accuracy of program generation drops sharply with progressive decoding. In this paper, we propose a non-autoregressive program generation framework, which facilitates program generation in parallel. Our framework, which independently generates complete program tuples containing both operators and operands, can significantly boost the speed of program generation while addressing the error accumulation issue. Our experiments on the MultiHiertt dataset shows that our model can bring about large improvements (+7.97 EM and +6.38 F1 points) over the strong baseline, establishing the new state-of-the-art performance, while being much faster (21x) in program generation. The performance drop of our method is also significantly smaller than the baseline with increasing numbers of numerical reasoning steps.
The presence of bias in deep models leads to unfair outcomes for certain demographic subgroups. Research in bias focuses primarily on facial recognition and attribute prediction with scarce emphasis on face detection. Existing studies consider face detection as binary classification into 'face' and 'non-face' classes. In this work, we investigate possible bias in the domain of face detection through facial region localization which is currently unexplored. Since facial region localization is an essential task for all face recognition pipelines, it is imperative to analyze the presence of such bias in popular deep models. Most existing face detection datasets lack suitable annotation for such analysis. Therefore, we web-curate the Fair Face Localization with Attributes (F2LA) dataset and manually annotate more than 10 attributes per face, including facial localization information. Utilizing the extensive annotations from F2LA, an experimental setup is designed to study the performance of four pre-trained face detectors. We observe (i) a high disparity in detection accuracies across gender and skin-tone, and (ii) interplay of confounding factors beyond demography. The F2LA data and associated annotations can be accessed at http://iab-rubric.org/index.php/F2LA.
Edge computing-based 3D perception has received attention in intelligent transportation systems (ITS) because real-time monitoring of traffic candidates potentially strengthens Vehicle-to-Everything (V2X) orchestration. Thanks to the capability of precisely measuring the depth information on surroundings from LiDAR, the increasing studies focus on lidar-based 3D detection, which significantly promotes the development of 3D perception. Few methods met the real-time requirement of edge deployment because of high computation-intensive operations. Moreover, an inconsistency problem of object detection remains uncovered in the pointcloud domain due to large sparsity. This paper thoroughly analyses this problem, comprehensively roused by recent works on determining inconsistency problems in the image specialisation. Therefore, we proposed a 3D harmonic loss function to relieve the pointcloud based inconsistent predictions. Moreover, the feasibility of 3D harmonic loss is demonstrated from a mathematical optimization perspective. The KITTI dataset and DAIR-V2X-I dataset are used for simulations, and our proposed method considerably improves the performance than benchmark models. Further, the simulative deployment on an edge device (Jetson Xavier TX) validates our proposed model's efficiency. Our code is open-source and publicly available.
The success of automatic speaker verification shows that discriminative speaker representations can be extracted from neutral speech. However, as a kind of non-verbal voice, laughter should also carry speaker information intuitively. Thus, this paper focuses on exploring speaker verification about utterances containing non-verbal laughter segments. We collect a set of clips with laughter components by conducting a laughter detection script on VoxCeleb and part of the CN-Celeb dataset. To further filter untrusted clips, probability scores are calculated by our binary laughter detection classifier, which is pre-trained by pure laughter and neutral speech. After that, based on the clips whose scores are over the threshold, we construct trials under two different evaluation scenarios: Laughter-Laughter (LL) and Speech-Laughter (SL). Then a novel method called Laughter-Splicing based Network (LSN) is proposed, which can significantly boost performance in both scenarios and maintain the performance on the neutral speech, such as the VoxCeleb1 test set. Specifically, our system achieves relative 20% and 22% improvement on Laughter-Laughter and Speech-Laughter trials, respectively. The meta-data and sample clips have been released at https://github.com/nevermoreLin/Laugh_LSN.
Scoring a large number of candidates precisely in several milliseconds is vital for industrial pre-ranking systems. Existing pre-ranking systems primarily adopt the \textbf{two-tower} model since the ``user-item decoupling architecture'' paradigm is able to balance the \textit{efficiency} and \textit{effectiveness}. However, the cost of high efficiency is the neglect of the potential information interaction between user and item towers, hindering the prediction accuracy critically. In this paper, we show it is possible to design a two-tower model that emphasizes both information interactions and inference efficiency. The proposed model, IntTower (short for \textit{Interaction enhanced Two-Tower}), consists of Light-SE, FE-Block and CIR modules. Specifically, lightweight Light-SE module is used to identify the importance of different features and obtain refined feature representations in each tower. FE-Block module performs fine-grained and early feature interactions to capture the interactive signals between user and item towers explicitly and CIR module leverages a contrastive interaction regularization to further enhance the interactions implicitly. Experimental results on three public datasets show that IntTower outperforms the SOTA pre-ranking models significantly and even achieves comparable performance in comparison with the ranking models. Moreover, we further verify the effectiveness of IntTower on a large-scale advertisement pre-ranking system. The code of IntTower is publicly available\footnote{https://github.com/archersama/IntTower}
Video anomaly detection aims to discover abnormal events in videos, and the principal objects are target objects such as people and vehicles. Each target in the video data has rich spatio-temporal context information. Most existing methods only focus on the temporal context, ignoring the role of the spatial context in anomaly detection. The spatial context information represents the relationship between the detection target and surrounding targets. Anomaly detection makes a lot of sense. To this end, a video anomaly detection algorithm based on target spatio-temporal context fusion is proposed. Firstly, the target in the video frame is extracted through the target detection network to reduce background interference. Then the optical flow map of two adjacent frames is calculated. Motion features are used multiple targets in the video frame to construct spatial context simultaneously, re-encoding the target appearance and motion features, and finally reconstructing the above features through the spatio-temporal dual-stream network, and using the reconstruction error to represent the abnormal score. The algorithm achieves frame-level AUCs of 98.5% and 86.3% on the UCSDped2 and Avenue datasets, respectively. On the UCSDped2 dataset, the spatio-temporal dual-stream network improves frames by 5.1% and 0.3%, respectively, compared to the temporal and spatial stream networks. After using spatial context encoding, the frame-level AUC is enhanced by 1%, which verifies the method's effectiveness.
Inferring the source information of greenhouse gases, such as methane, from spatially sparse sensor observations is an essential element in mitigating climate change. While it is well understood that the complex behavior of the atmospheric dispersion of such pollutants is governed by the Advection-Diffusion equation, it is difficult to directly apply the governing equations to identify the source location and magnitude (inverse problem) because of the spatially sparse and noisy observations, i.e., the pollution concentration is known only at the sensor locations and sensors sensitivity is limited. Here, we develop a multi-task learning framework that can provide high-fidelity reconstruction of the concentration field and identify emission characteristics of the pollution sources such as their location, emission strength, etc. from sparse sensor observations. We demonstrate that our proposed framework is able to achieve accurate reconstruction of the methane concentrations from sparse sensor measurements as well as precisely pin-point the location and emission strength of these pollution sources.
The optimization of high-dimensional black-box functions is a challenging problem. When a low-dimensional linear embedding structure can be assumed, existing Bayesian optimization (BO) methods often transform the original problem into optimization in a low-dimensional space. They exploit the low-dimensional structure and reduce the computational burden. However, we reveal that this approach could be limited or inefficient in exploring the high-dimensional space mainly due to the biased reconstruction of the high-dimensional queries from the low-dimensional queries. In this paper, we investigate a simple alternative approach: tackling the problem in the original high-dimensional space using the information from the learned low-dimensional structure. We provide a theoretical analysis of the exploration ability. Furthermore, we show that our method is applicable to batch optimization problems with thousands of dimensions without any computational difficulty. We demonstrate the effectiveness of our method on high-dimensional benchmarks and a real-world function.