Contrastive Language-Image pre-training (CLIP) learns rich representations via readily available supervisions of natural language. It could improve general performance on downstream vision tasks, including but not limited to zero-shot, long tail, segmentation, retrieval, caption and video. However, to the best of our knowledge, the visual interpretability of CLIP has not been studied yet. To provide visual explanations of its predictions, we propose the Image-Text Similarity Map (ITSM). Based on it, we surprisingly find that CLIP prefers the background regions than the foregrounds, and presenting erroneous visualization against human understanding. Experimentally, we find the devil is in the pooling part, where inappropriate pooling methods lead to a phenomenon called semantic shift. To correct and boost the visualization results, we propose the Masked Max Pooling, with attention map from the self-supervised image encoder. Meanwhile, interpretability task and recognition task require different representations. To address the problem, we propose the dual projections to cater this requirement. We integrate above methods as Interpretable Contrastive Language-Image pre-training (ICLIP). And experiments suggest ICLIP greatly improves the interpretability. For example, the nontrivial improvements are $32.85\%$ and $49.10\%$, respectively, on VOC 2012 dataset.
The traditional homography estimation pipeline consists of four main steps: feature detection, feature matching, outlier removal and transformation estimation. Recent deep learning models intend to address the homography estimation problem using a single convolutional network. While these models are trained in an end-to-end fashion to simplify the homography estimation problem, they lack the feature matching step and/or the outlier removal step, which are important steps in the traditional homography estimation pipeline. In this paper, we attempt to build a deep learning model that mimics all four steps in the traditional homography estimation pipeline. In particular, the feature matching step is implemented using the cost volume technique. To remove outliers in the cost volume, we treat this outlier removal problem as a denoising problem and propose a novel self-supervised loss to solve the problem. Extensive experiments on synthetic and real datasets demonstrate that the proposed model outperforms existing deep learning models.
While category-level 9DoF object pose estimation has emerged recently, previous correspondence-based or direct regression methods are both limited in accuracy due to the huge intra-category variances in object shape and color, etc. Orthogonal to them, this work presents a category-level object pose and size refiner CATRE, which is able to iteratively enhance pose estimate from point clouds to produce accurate results. Given an initial pose estimate, CATRE predicts a relative transformation between the initial pose and ground truth by means of aligning the partially observed point cloud and an abstract shape prior. In specific, we propose a novel disentangled architecture being aware of the inherent distinctions between rotation and translation/size estimation. Extensive experiments show that our approach remarkably outperforms state-of-the-art methods on REAL275, CAMERA25, and LM benchmarks up to a speed of ~85.32Hz, and achieves competitive results on category-level tracking. We further demonstrate that CATRE can perform pose refinement on unseen category. Code and trained models are available.
Active regression considers a linear regression problem where the learner receives a large number of data points but can only observe a small number of labels. Since online algorithms can deal with incremental training data and take advantage of low computational cost, we consider an online extension of the active regression problem: the learner receives data points one by one and immediately decides whether it should collect the corresponding labels. The goal is to efficiently maintain the regression of received data points with a small budget of label queries. We propose novel algorithms for this problem under $\ell_p$ loss where $p\in[1,2]$. To achieve a $(1+\epsilon)$-approximate solution, our proposed algorithms only require $\tilde{\mathcal{O}}(\epsilon^{-2} d \log(n\kappa))$ queries of labels, where $n$ is the number of data points and $\kappa$ is a quantity, called the condition number, of the data points. The numerical results verify our theoretical results and show that our methods have comparable performance with offline active regression algorithms.
Developing an AI-assisted gland segmentation method from histology images is critical for automatic cancer diagnosis and prognosis; however, the high cost of pixel-level annotations hinders its applications to broader diseases. Existing weakly-supervised semantic segmentation methods in computer vision achieve degenerative results for gland segmentation, since the characteristics and problems of glandular datasets are different from general object datasets. We observe that, unlike natural images, the key problem with histology images is the confusion of classes owning to morphological homogeneity and low color contrast among different tissues. To this end, we propose a novel method Online Easy Example Mining (OEEM) that encourages the network to focus on credible supervision signals rather than noisy signals, therefore mitigating the influence of inevitable false predictions in pseudo-masks. According to the characteristics of glandular datasets, we design a strong framework for gland segmentation. Our results exceed many fully-supervised methods and weakly-supervised methods for gland segmentation over 4.4% and 6.04% at mIoU, respectively. Code is available at https://github.com/xmed-lab/OEEM.
Self-supervised learning (SSL) achieves great success in monaural speech enhancement, while the accuracy of the target speech estimation, particularly for unseen speakers, remains inadequate with existing pre-tasks. As speech signal contains multi-faceted information including speaker identity, paralinguistics, and spoken content, the latent representation for speech enhancement becomes a tough task. In this paper, we study the effectiveness of each feature which is commonly used in speech enhancement and exploit the feature combination in the SSL case. Besides, we propose an ensemble training strategy. The latent representation of the clean speech signal is learned, meanwhile, the dereverberated mask and the estimated ratio mask are exploited to denoise and dereverberate the mixture. The latent representation learning and the masks estimation are considered as two pre-tasks in the training stage. In addition, to study the effectiveness between the pre-tasks, we compare different training routines to train the model and further refine the performance. The NOISEX and DAPS corpora are used to evaluate the efficacy of the proposed method, which also outperforms the state-of-the-art methods.
Designing better machine translation systems by considering auxiliary inputs such as images has attracted much attention in recent years. While existing methods show promising performance over the conventional text-only translation systems, they typically require paired text and image as input during inference, which limits their applicability to real-world scenarios. In this paper, we introduce a visual hallucination framework, called VALHALLA, which requires only source sentences at inference time and instead uses hallucinated visual representations for multimodal machine translation. In particular, given a source sentence an autoregressive hallucination transformer is used to predict a discrete visual representation from the input text, and the combined text and hallucinated representations are utilized to obtain the target translation. We train the hallucination transformer jointly with the translation transformer using standard backpropagation with cross-entropy losses while being guided by an additional loss that encourages consistency between predictions using either ground-truth or hallucinated visual representations. Extensive experiments on three standard translation datasets with a diverse set of language pairs demonstrate the effectiveness of our approach over both text-only baselines and state-of-the-art methods. Project page: http://www.svcl.ucsd.edu/projects/valhalla.
Despite achieving superior performance in human-level control problems, unlike humans, deep reinforcement learning (DRL) lacks high-order intelligence (e.g., logic deduction and reuse), thus it behaves ineffectively than humans regarding learning and generalization in complex problems. Previous works attempt to directly synthesize a white-box logic program as the DRL policy, manifesting logic-driven behaviors. However, most synthesis methods are built on imperative or declarative programming, and each has a distinct limitation, respectively. The former ignores the cause-effect logic during synthesis, resulting in low generalizability across tasks. The latter is strictly proof-based, thus failing to synthesize programs with complex hierarchical logic. In this paper, we combine the above two paradigms together and propose a novel Generalizable Logic Synthesis (GALOIS) framework to synthesize hierarchical and strict cause-effect logic programs. GALOIS leverages the program sketch and defines a new sketch-based hybrid program language for guiding the synthesis. Based on that, GALOIS proposes a sketch-based program synthesis method to automatically generate white-box programs with generalizable and interpretable cause-effect logic. Extensive evaluations on various decision-making tasks with complex logic demonstrate the superiority of GALOIS over mainstream baselines regarding the asymptotic performance, generalizability, and great knowledge reusability across different environments.
As an essential prerequisite for developing a medical intelligent assistant system, medical image segmentation has received extensive research and concentration from the neural network community. A series of UNet-like networks with encoder-decoder architecture has achieved extraordinary success, in which UNet2+ and UNet3+ redesign skip connections, respectively proposing dense skip connection and full-scale skip connection and dramatically improving compared with UNet in medical image segmentation. However, UNet2+ lacks sufficient information explored from the full scale, which will affect the learning of organs' location and boundary. Although UNet3+ can obtain the full-scale aggregation feature map, owing to the small number of neurons in the structure, it does not satisfy the segmentation of tiny objects when the number of samples is small. This paper proposes a novel network structure combining dense skip connections and full-scale skip connections, named UNet-sharp (UNet\#) for its shape similar to symbol \#. The proposed UNet\# can aggregate feature maps of different scales in the decoder sub-network and capture fine-grained details and coarse-grained semantics from the full scale, which benefits learning the exact location and accurately segmenting the boundary of organs or lesions. We perform deep supervision for model pruning to speed up testing and make it possible for the model to run on mobile devices; furthermore, designing two classification-guided modules to reduce false positives achieves more accurate segmentation results. Various experiments of semantic segmentation and instance segmentation on different modalities (EM, CT, MRI) and dimensions (2D, 3D) datasets, including the nuclei, brain tumor, liver, and lung, demonstrate that the proposed method outperforms state-of-the-art models.
Gaussian graphical regression is a powerful means that regresses the precision matrix of a Gaussian graphical model on covariates, permitting the numbers of the response variables and covariates to far exceed the sample size. Model fitting is typically carried out via separate node-wise lasso regressions, ignoring the network-induced structure among these regressions. Consequently, the error rate is high, especially when the number of nodes is large. We propose a multi-task learning estimator for fitting Gaussian graphical regression models; we design a cross-task group sparsity penalty and a within task element-wise sparsity penalty, which govern the sparsity of active covariates and their effects on the graph, respectively. For computation, we consider an efficient augmented Lagrangian algorithm, which solves subproblems with a semi-smooth Newton method. For theory, we show that the error rate of the multi-task learning based estimates has much improvement over that of the separate node-wise lasso estimates, because the cross-task penalty borrows information across tasks. To address the main challenge that the tasks are entangled in a complicated correlation structure, we establish a new tail probability bound for correlated heavy-tailed (sub-exponential) variables with an arbitrary correlation structure, a useful theoretical result in its own right. Finally, the utility of our method is demonstrated through simulations as well as an application to a gene co-expression network study with brain cancer patients.