Large language models (LLMs) are powerful tools for content moderation, but their inference costs and latency make them prohibitive for casual use on large datasets, such as the Google Ads repository. This study proposes a method for scaling up LLM reviews for content moderation in Google Ads. First, we use heuristics to select candidates via filtering and duplicate removal, and create clusters of ads for which we select one representative ad per cluster. We then use LLMs to review only the representative ads. Finally, we propagate the LLM decisions for the representative ads back to their clusters. This method reduces the number of reviews by more than 3 orders of magnitude while achieving a 2x recall compared to a baseline non-LLM model. The success of this approach is a strong function of the representations used in clustering and label propagation; we found that cross-modal similarity representations yield better results than uni-modal representations.

Many vision-based applications rely on logistic regression for embedding classification within a probabilistic context, such as recognition in images and videos or identifying disease-specific image phenotypes from neuroimages. Logistic regression, however, often performs poorly when trained on data that is noisy, has irrelevant features, or when the samples are distributed across the classes in an imbalanced setting; a common occurrence in visual recognition tasks. To deal with those issues, researchers generally rely on adhoc regularization techniques or model a subset of these issues. We instead propose a mathematically sound logistic regression model that selects a subset of (relevant) features and (informative and balanced) set of samples during the training process. The model does so by applying cardinality constraints (via ℓ0-`norm' sparsity) on the features and samples. ℓ0defines sparsity in mathematical settings but in practice has mostly been approximated (e.g., via ℓ1or its variations) for computational simplicity. We prove that a local minimum to the non-convex optimization problems induced by cardinality constraints can be computed by combining block coordinate descent with penalty decomposition. On synthetic, image recognition, and neuroimaging datasets, we show that the accuracy of the method is higher than alternative methods and classifiers commonly used in the literature.

We propose a new method for deformable registration of pre-operative and post-recurrence brain MR scans of glioma patients. Performing this type of intra-subject registration is challenging as tumor, resection, recurrence, and edema cause large deformations, missing correspondences, and inconsistent intensity profiles between the scans. To address this challenging task, our method, called PORTR, explicitly accounts for pathological information. It segments tumor, resection cavity, and recurrence based on models specific to each scan. PORTR then uses the resulting maps to exclude pathological regions from the image-based correspondence term while simultaneously measuring the overlap between the aligned tumor and resection cavity. Embedded into a symmetric registration framework, we determine the optimal solution by taking advantage of both discrete and continuous search methods. We apply our method to scans of 24 glioma patients. Both quantitative and qualitative analysis of the results clearly show that our method is superior to other state-of-the-art approaches.

We present a new energy model for optical flow estimation on discrete MRF framework. The proposed model yields discrete analog to the prevailing model with diffusion tensor-based regularizer, which has been optimized by variational approach. Inspired from the fact that the regularization process works as a convolution kernel filtering, we formulate the difference between original flow and filtered flow as a smoothness prior. Then the discrete framework enables us to employ a robust penalizer less concerning convexity and differentiability of the energy function. In addition, we provide a new kernel design based on the bilateral filter, adaptively controlling intensity variance according to the local statistics. The proposed kernel simultaneously addresses over-segmentation and over-smoothing problems, which is hard to achieve by tuning parameters. Involving a complex graph structure with large label sets, this work also presents a strategy to efficiently reduce memory requirement and computational time to a tolerable state. Experimental result shows the proposed method yields plausible results on the various data sets including large displacement and textured region.

This paper proposes a new rolled fingerprint construction approach incorporating a state-of-the-art nonrigid image registration method based upon a Markov random field (MRF) energy model. The proposed method finds dense correspondences between images from a rolled fingerprint sequence and warps the entire fingerprint area to synthesize a rolled fingerprint. This method can generate conceptually more accurate rolled fingerprints by preserving the geometric properties of the finger surface as opposed to ink-based rolled impressions and other existing rolled fingerprint construction methods. To verify the accuracy of the proposed method, various comparative experiments were designed to reveal differences among the rolled construction methods. The results show that the proposed method is significantly superior in various aspects compared to previous approaches.

This paper presents an efficient feature-based nonrigid registration method for multiphase liver CT volumes. While radiologists routinely examine multiphase liver CT to detect hepatic diseases, they usually search corresponding points between 3D CT volumes by visual inspections using 2D slice images. As the liver is a deformable organ, there exist complex nonrigid transformations between liver CT volumes obtained at difference time points (phases). We introduce a fully automatic registration application for multiphase liver CT volumes. For two given liver CT volumes, we extract 3D features with their descriptors, and estimate correspondences by finding nearest neighbor in descriptor space. An energy function is constructed using the correspondence information and the smoothness measure of free-form deformation model based on B-splines. We integrate an approximated smoothness energy function and a robust correspondence energy estimator controlled by the confidence radius of the matching distance in this energy model. The energy function is optimized by sequentially reducing the confidence radius, and outlier correspondences are discarded systematically during convergence. We propose a highly efficient optimization procedure using the preconditioned nonlinear conjugate gradient method. In the experiments, we will provide quantitative and qualitative results on synthetic and clinical data sets. 1

An efficient registration algorithm working on non-rigid 3D objects is presented. We formulate the registration as a discrete labeling problem on MRFs model whose energy can be minimized by optimization techniques in the literature. Due to the huge search range in three-dimensional space, previous approaches produces a vast amount of labels for a node in the MRFs graph. To reduce the number of labels, we decompose a node into three nodes so that the labels in each node represent just one-dimensional displacement. This procedure introduces a factor node with a clique potential of size three, defining ternary interaction between the decomposed nodes. We convert the factor node into pairwise interactions and adopt the tree-reweighted message passing technique, which guarantees the convergence of lower bound of the energy function. In experiments we use clinical and synthetically deformed 3D medical images. Result shows the proposed method enhances computational efficiency without loss of accuracy. 1

This paper presents a robust warping method for minutiae based fingerprint matching approaches. In this method, a deformable fingerprint surface is described using a triangular mesh model. For given two extracted minutiae sets and their correspondences, the proposed method constructs an energy function using a robust correspondence energy estimator and smoothness measuring of the mesh model. We obtain a convergent deformation pattern using an efficient gradient based energy optimization method. This energy optimization approach deals successfully with deformation errors caused by outliers, which are more difficult problems for the thin-plate spline (TPS) model. The proposed method is fast and the run-time performance is comparable with the method based on the TPS model. In the experiments, we provide a visual inspection of warping results on given correspondences and quantitative results using database.