The transformer is a widely-used building block in modern neural networks. However, when applied to audio data, the transformer's acausal behaviour, which we term Acausal Attention (AA), has generally limited its application to offline tasks. In this paper we introduce Streaming Attention (SA), which operates causally with fixed latency, and requires lower compute and memory resources than AA to train. Next, we introduce Low Latency Streaming Attention (LLSA), a method which combines multiple SA layers without latency build-up proportional to the layer count. Comparative analysis between AA, SA and LLSA on Automatic Speech Recognition (ASR) and Speech Emotion Recognition (SER) tasks are presented. The results show that causal SA-based networks with fixed latencies of a few seconds (e.g. 1.8 seconds) and LLSA networks with latencies as short as 300 ms can perform comparably with acausal (AA) networks. We conclude that SA and LLSA methods retain many of the benefits of conventional acausal transformers, but with latency characteristics that make them practical to run in real-time streaming applications.
A new class of iterated linearization-based nonlinear filters, dubbed dynamically iterated filters, is presented. Contrary to regular iterated filters such as the iterated extended Kalman filter (IEKF), iterated unscented Kalman filter (IUKF) and iterated posterior linearization filter (IPLF), dynamically iterated filters also take nonlinearities in the transition model into account. The general filtering algorithm is shown to essentially be a (locally over one time step) iterated Rauch-Tung-Striebel smoother. Three distinct versions of the dynamically iterated filters are especially investigated: analogues to the IEKF, IUKF and IPLF. The developed algorithms are evaluated on 25 different noise configurations of a tracking problem with a nonlinear transition model and linear measurement model, a scenario where conventional iterated filters are not useful. Even in this "simple" scenario, the dynamically iterated filters are shown to have superior root mean-squared error performance as compared to their respective baselines, the EKF and UKF. Particularly, even though the EKF diverges in 22 out of 25 configurations, the dynamically iterated EKF remains stable in 20 out of 25 scenarios, only diverging under high noise.
Recommendation systems are a core feature of social media companies with their uses including recommending organic and promoted contents. Many modern recommendation systems are split into multiple stages - candidate generation and heavy ranking - to balance computational cost against recommendation quality. We focus on the candidate generation phase of a large-scale ads recommendation problem in this paper, and present a machine learning first heterogeneous re-architecture of this stage which we term TwERC. We show that a system that combines a real-time light ranker with sourcing strategies capable of capturing additional information provides validated gains. We present two strategies. The first strategy uses a notion of similarity in the interaction graph, while the second strategy caches previous scores from the ranking stage. The graph based strategy achieves a 4.08% revenue gain and the rankscore based strategy achieves a 1.38% gain. These two strategies have biases that complement both the light ranker and one another. Finally, we describe a set of metrics that we believe are valuable as a means of understanding the complex product trade offs inherent in industrial candidate generation systems.
We develop new tools in the theory of nonlinear random matrices and apply them to study the performance of the Sum of Squares (SoS) hierarchy on average-case problems. The SoS hierarchy is a powerful optimization technique that has achieved tremendous success for various problems in combinatorial optimization, robust statistics and machine learning. It's a family of convex relaxations that lets us smoothly trade off running time for approximation guarantees. In recent works, it's been shown to be extremely useful for recovering structure in high dimensional noisy data. It also remains our best approach towards refuting the notorious Unique Games Conjecture. In this work, we analyze the performance of the SoS hierarchy on fundamental problems stemming from statistics, theoretical computer science and statistical physics. In particular, we show subexponential-time SoS lower bounds for the problems of the Sherrington-Kirkpatrick Hamiltonian, Planted Slightly Denser Subgraph, Tensor Principal Components Analysis and Sparse Principal Components Analysis. These SoS lower bounds involve analyzing large random matrices, wherein lie our main contributions. These results offer strong evidence for the truth of and insight into the low-degree likelihood ratio hypothesis, an important conjecture that predicts the power of bounded-time algorithms for hypothesis testing. We also develop general-purpose tools for analyzing the behavior of random matrices which are functions of independent random variables. Towards this, we build on and generalize the matrix variant of the Efron-Stein inequalities. In particular, our general theorem on matrix concentration recovers various results that have appeared in the literature. We expect these random matrix theory ideas to have other significant applications.
We propose a novel solution for the task of video panoptic segmentation, that simultaneously predicts pixel-level semantic and instance segmentation and generates clip-level instance tracks. Our network, named VPS-Transformer, with a hybrid architecture based on the state-of-the-art panoptic segmentation network Panoptic-DeepLab, combines a convolutional architecture for single-frame panoptic segmentation and a novel video module based on an instantiation of the pure Transformer block. The Transformer, equipped with attention mechanisms, models spatio-temporal relations between backbone output features of current and past frames for more accurate and consistent panoptic estimates. As the pure Transformer block introduces large computation overhead when processing high resolution images, we propose a few design changes for a more efficient compute. We study how to aggregate information more effectively over the space-time volume and we compare several variants of the Transformer block with different attention schemes. Extensive experiments on the Cityscapes-VPS dataset demonstrate that our best model improves the temporal consistency and video panoptic quality by a margin of 2.2%, with little extra computation.
We focus on decentralized stochastic non-convex optimization, where $n$ agents work together to optimize a composite objective function which is a sum of a smooth term and a non-smooth convex term. To solve this problem, we propose two single-time scale algorithms: Prox-DASA and Prox-DASA-GT. These algorithms can find $\epsilon$-stationary points in $\mathcal{O}(n^{-1}\epsilon^{-2})$ iterations using constant batch sizes (i.e., $\mathcal{O}(1)$). Unlike prior work, our algorithms achieve a comparable complexity result without requiring large batch sizes, more complex per-iteration operations (such as double loops), or stronger assumptions. Our theoretical findings are supported by extensive numerical experiments, which demonstrate the superiority of our algorithms over previous approaches.
Bayesian Optimization (BO) is a class of black-box, surrogate-based heuristics that can efficiently optimize problems that are expensive to evaluate, and hence admit only small evaluation budgets. BO is particularly popular for solving numerical optimization problems in industry, where the evaluation of objective functions often relies on time-consuming simulations or physical experiments. However, many industrial problems depend on a large number of parameters. This poses a challenge for BO algorithms, whose performance is often reported to suffer when the dimension grows beyond 15 variables. Although many new algorithms have been proposed to address this problem, it is not well understood which one is the best for which optimization scenario. In this work, we compare five state-of-the-art high-dimensional BO algorithms, with vanilla BO and CMA-ES on the 24 BBOB functions of the COCO environment at increasing dimensionality, ranging from 10 to 60 variables. Our results confirm the superiority of BO over CMA-ES for limited evaluation budgets and suggest that the most promising approach to improve BO is the use of trust regions. However, we also observe significant performance differences for different function landscapes and budget exploitation phases, indicating improvement potential, e.g., through hybridization of algorithmic components.
Introduced by Hinton et al. in 2012, dropout has stood the test of time as a regularizer for preventing overfitting in neural networks. In this study, we demonstrate that dropout can also mitigate underfitting when used at the start of training. During the early phase, we find dropout reduces the directional variance of gradients across mini-batches and helps align the mini-batch gradients with the entire dataset's gradient. This helps counteract the stochasticity of SGD and limit the influence of individual batches on model training. Our findings lead us to a solution for improving performance in underfitting models - early dropout: dropout is applied only during the initial phases of training, and turned off afterwards. Models equipped with early dropout achieve lower final training loss compared to their counterparts without dropout. Additionally, we explore a symmetric technique for regularizing overfitting models - late dropout, where dropout is not used in the early iterations and is only activated later in training. Experiments on ImageNet and various vision tasks demonstrate that our methods consistently improve generalization accuracy. Our results encourage more research on understanding regularization in deep learning and our methods can be useful tools for future neural network training, especially in the era of large data. Code is available at https://github.com/facebookresearch/dropout .
Dressed people reconstruction from images is a popular task with promising applications in the creative media and game industry. However, most existing methods reconstruct the human body and garments as a whole with the supervision of 3D models, which hinders the downstream interaction tasks and requires hard-to-obtain data. To address these issues, we propose an unsupervised separated 3D garments and human reconstruction model (USR), which reconstructs the human body and authentic textured clothes in layers without 3D models. More specifically, our method proposes a generalized surface-aware neural radiance field to learn the mapping between sparse multi-view images and geometries of the dressed people. Based on the full geometry, we introduce a Semantic and Confidence Guided Separation strategy (SCGS) to detect, segment, and reconstruct the clothes layer, leveraging the consistency between 2D semantic and 3D geometry. Moreover, we propose a Geometry Fine-tune Module to smooth edges. Extensive experiments on our dataset show that comparing with state-of-the-art methods, USR achieves improvements on both geometry and appearance reconstruction while supporting generalizing to unseen people in real time. Besides, we also introduce SMPL-D model to show the benefit of the separated modeling of clothes and the human body that allows swapping clothes and virtual try-on.
Retrieval with extremely long queries and documents is a well-known and challenging task in information retrieval and is commonly known as Query-by-Document (QBD) retrieval. Specifically designed Transformer models that can handle long input sequences have not shown high effectiveness in QBD tasks in previous work. We propose a Re-Ranker based on the novel Proportional Relevance Score (RPRS) to compute the relevance score between a query and the top-k candidate documents. Our extensive evaluation shows RPRS obtains significantly better results than the state-of-the-art models on five different datasets. Furthermore, RPRS is highly efficient since all documents can be pre-processed, embedded, and indexed before query time which gives our re-ranker the advantage of having a complexity of O(N) where N is the total number of sentences in the query and candidate documents. Furthermore, our method solves the problem of the low-resource training in QBD retrieval tasks as it does not need large amounts of training data, and has only three parameters with a limited range that can be optimized with a grid search even if a small amount of labeled data is available. Our detailed analysis shows that RPRS benefits from covering the full length of candidate documents and queries.