Knowledge Distillation (KD) is a powerful approach for compressing a large model into a smaller, more efficient model, particularly beneficial for latency-sensitive applications like recommender systems. However, current KD research predominantly focuses on Computer Vision (CV) and NLP tasks, overlooking unique data characteristics and challenges inherent to recommender systems. This paper addresses these overlooked challenges, specifically: (1) mitigating data distribution shifts between teacher and student models, (2) efficiently identifying optimal teacher configurations within time and budgetary constraints, and (3) enabling computationally efficient and rapid sharing of teacher labels to support multiple students. We present a robust KD system developed and rigorously evaluated on multiple large-scale personalized video recommendation systems within Google. Our live experiment results demonstrate significant improvements in student model performance while ensuring consistent and reliable generation of high-quality teacher labels from a continuous data stream of data.

In this paper, we introduce a large scale multi-objective ranking system for recommending what video to watch next on an industrial video sharing platform. The system faces many real-world challenges, including the presence of multiple competing ranking objectives, as well as implicit selection biases in user feedback. To tackle these challenges, we explored a variety of soft-parameter sharing techniques such as Multi-gate Mixture-of-Experts so as to efficiently optimize for multiple ranking objectives. Additionally, we mitigated the selection biases by adopting a Wide & Deep framework. We demonstrated that our proposed techniques can lead to substantial improvements on recommendation quality on one of the world's largest video sharing platforms.

We present a novel probabilistic shape representation that implicitly includes prior anatomical volume and adjacency information, termed the generalized log-ratio (GLR) representation. We demonstrate the usefulness of this representation in the task of thigh muscle segmentation. Analysis of the shapes and sizes of thigh muscles can lead to a better understanding of the effects of chronic obstructive pulmonary disease (COPD), which often results in skeletal muscle weakness in lower limbs. However, segmenting these muscles from one another is difficult due to a lack of distinctive features and inter-muscular boundaries that are difficult to detect. We overcome these difficulties by building a shape model in the space of GLR representations. We remove pose variability from the model by employing a presegmentation-based alignment scheme. We also design a rotationally invariant random forest boundary detector that learns common appearances of the interface between muscles from training data. We combine the shape model and the boundary detector into a fully automatic globally optimal segmentation technique. Our segmentation technique produces a probabilistic segmentation that can be used to generate uncertainty information, which can be used to aid subsequent analysis. Our experiments on challenging 3D magnetic resonance imaging data sets show that the use of the GLR representation improves the segmentation accuracy, and yields an average Dice similarity coefficient of 0.808 ±0.074, comparable to other state-of-the-art thigh segmentation techniques.

Sources of uncertainty in the boundaries of structures in medical images have motivated the use of probabilistic labels in segmentation applications. An important component in many medical image segmentation tasks is the use of a shape model, often generated by applying statistical techniques to training data. Standard statistical techniques (e.g., principal component analysis) often assume data lies in an unconstrained vector space, but probabilistic labels are constrained to the unit simplex. If these statistical techniques are used directly on probabilistic labels, relative uncertainty information can be sacrificed. A standard method for facilitating analysis of probabilistic labels is to map them to a vector space using the LogOdds transform. However, the LogOdds transform is asymmetric in one of the labels, which skews results in some applications. The isometric log-ratio (ILR) transform is a symmetrized version of the LogOdds transform, and is so named as it is an isometry between the Aitchison geometry, the inherent geometry of the simplex, and standard Euclidean geometry. We explore how to interpret the Aitchison geometry when applied to probabilistic labels in medical image segmentation applications. We demonstrate the differences when applying the LogOdds transform or the ILR transform to probabilistic labels prior to statistical analysis. Specifically, we show that statistical analysis of ILR transformed data better captures the variability of anatomical shapes in cases where multiple different foreground regions share boundaries (as opposed to foreground-background boundaries).

The inclusion of shape and appearance priors have proven useful for obtaining more accurate and plausible segmentations, especially for complex objects with multiple parts. In this paper, we augment the popular Mum ford-Shah model to incorporate two important geometrical constraints, termed containment and detachment, between different regions with a specified minimum distance between their boundaries. Our method is able to handle multiple instances of multi-part objects defined by these geometrical constraints using a single labeling function while maintaining global optimality. We demonstrate the utility and advantages of these two constraints and show that the proposed convex continuous method is superior to other state-of-the-art methods, including its discrete counterpart, in terms of memory usage, and metrication errors.

Image segmentation is often performed via the minimization of an energy function over a domain of possible segmentations. The effectiveness and applicability of such methods depends greatly on the properties of the energy function and its domain, and on what information can be encoded by it. Here we propose an energy function that achieves several important goals. Specifically, our energy function is convex and incorporates shape prior information while simultaneously generating a probabilistic segmentation for multiple regions. Our energy function represents multi-region probabilistic segmentations as elements of a vector space using the isometric log-ratio (ILR) transformation. To our knowledge, these four goals (convex, with shape priors, multi-region, and probabilistic) do not exist together in any other method, and this is the first time ILR is used in an image segmentation method. We provide examples demonstrating the usefulness of these features.